Grothendieck’s Theorem, past and present

by

Gilles Pisier∗

Texas A&M University

College Station, TX 77843, U. S. A.

and

Universite Paris VI

Equipe d’Analyse, Case 186, 75252

Paris Cedex 05, France

January 26, 2011

Abstract

Probably the most famous of Grothendieck’s contributions to Banach space theory is theresult that he himself described as “the fundamental theorem in the metric theory of tensorproducts”. That is now commonly referred to as “Grothendieck’s theorem” (GT in short), orsometimes as “Grothendieck’s inequality”. This had a major impact first in Banach space theory(roughly after 1968), then, later on, in C∗-algebra theory, (roughly after 1978). More recently,in this millennium, a new version of GT has been successfully developed in the framework of“operator spaces” or non-commutative Banach spaces. In addition, GT independently surfacedin several quite unrelated fields: in connection with Bell’s inequality in quantum mechanics, ingraph theory where the Grothendieck constant of a graph has been introduced and in computerscience where the Grothendieck inequality is invoked to replace certain NP hard problems byothers that can be treated by “semidefinite programming’ and hence solved in polynomial time.In this expository paper, we present a review of all these topics, starting from the original GT.We concentrate on the more recent developments and merely outline those of the first Banachspace period since detailed accounts of that are already available, for instance the author’s 1986CBMS notes.

∗Partially supported by NSF grant 0503688

Contents

1 Introduction 1

2 Classical GT 7

3 The Grothendieck constants 15

4 The Hahn-Banach argument 17

5 The “little” GT 20

6 Banach spaces satisfying GT and operator ideals 22

7 Non-commutative GT 23

8 Non-commutative “little GT” 25

9 Non-commutative Khintchine inequality 26

10 Maurey factorization 32

11 Best constants (Non-commutative case) 33

12 C∗C∗C∗-algebra tensor products, Nuclearity 35

13 Operator spaces, c.b. maps, Minimal tensor product 40

14 Haagerup tensor product 46

15 The operator Hilbert space OH 48

16 GT and random matrices 49

17 GT for exact operator spaces 54

18 GT for operator spaces 56

19 GT and Quantum mechanics: EPR and Bell’s inequality 63

20 Trilinear counterexamples 66

21 Some open problems 69

22 GT in graph theory and computer science 70

1. Introduction

The Resume saga In 1953, Grothendieck published an extraordinary paper [36] entitled “Resumede la theorie metrique des produits tensoriels topologiques,” now often jokingly referred to as“Grothendieck’s resume”(!). Just like his thesis ([38]), this was devoted to tensor products oftopological vector spaces, but in sharp contrast with the thesis devoted to the locally convex case,the “Resume” was exclusively concerned with Banach spaces (“theorie metrique”). The centralresult of this long paper (“Theoreme fondamental de la theorie metrique des produits tensorielstopologiques”) is now called Grothendieck’s Theorem (or Grothendieck’s inequality). We will referto it as GT. Informally, one could describe GT as a surprising and non-trivial relation betweenHilbert space (e.g. L2) and the two fundamental Banach spaces L∞, L1 (here L∞ can be replaced bythe space C(S) of continuous functions on a compact set S). That relationship was expressed by aninequality involving 3 fundamental tensor norms that are described at the end of this introduction.(Here the reader is advised to jump below to read both the original formulation of GT in (1.14) andan elementary one-not using tensor products-in Theorem 2.4). The paper went on to investigatethe 14 other tensor norms that can be derived from the first 3. When it appeared this astonishingpaper was virtually ignored.

Although the paper was reviewed in Math Reviews by none less than Dvoretzky, it seems tohave been widely ignored, until Lindenstrauss and PeÃlczynski’s 1968 paper [89] drew attention toit. Many explanations come to mind: it was written in French, published in a Brazilian journalwith very limited circulation and, in a major reversal from the author’s celebrated thesis, it ignoredlocally convex questions and concentrated exclusively on Banach spaces, a move that probably wentagainst the tide at that time.

The situation changed radically (15 years later) when Lindenstrauss and PeÃlczynski [89] discov-ered the paper’s numerous gems, including the solutions to several open questions that had beenraised after its appearance! Alongside with [89], Pietsch’s theory of p-absolutely summing opera-tors with its celebrated factorization had just appeared and it contributed to the resurgence of the“resume” although Pietsch’s emphasis was on ideals of operators (i.e. duals) rather than on ten-sor products (i.e. preduals), see [111]. Lindenstrauss and PeÃlczynski’s beautiful paper completelydissected the “resume”, rewriting the proof of the fundamental theorem (i.e. GT) and giving of itmany reformulations, some very elementary ones (see Theorem 2.4 below) as well as some morerefined consequences involving absolutely summing operators between Lp-spaces, a generalized no-tion of Lp-spaces that had just been introduced by Lindenstrauss and Rosenthal. Their work alsoemphasized a very useful factorization of operators from a L∞ space to a Hilbert space (cf. also[27]) that is much easier to prove that GT itself, and is now commonly called the “little GT” (see§ 5).

Despite all these efforts, the only known proof of GT remained the original one until Maurey[97] found the first new proof using an extrapolation method that turned out to be extremely fruit-ful. After that, several new proofs were given, notably a strikingly short one based on HarmonicAnalysis by PeÃlczynski and Wojtaszczyk (see [115, p. 68]). Moreover Krivine [84, 85] managed toimprove the original proof and the bound for the Grothendieck constant KG, that remains the bestup to now. Both Krivine’s and the original proof of GT are included in § 2 below.

More recent results The “Resume” ended with a remarkable list of six problems, on top ofwhich was the famous “Approximation problem” solved by Enflo in 1972. By 1981, all of the otherproblems had been solved (except for the value of the constant KG). Our CBMS notes from 1986[115] contain a detailed account of the work from that period, so we will not expand on that here.We merely summarize this very briefly in § 6 below. In the present survey, we choose to focus

1

solely on GT. One of the six problems was to prove a non-commutative version of GT for boundedbilinear forms on C∗-algebras. This was proved in [112] with some restriction and in [45] in fullgenerality. Both proofs use a certain form of non-commutative Khintchine inequality in L4. Thenon-commutative GT (see § 7), or actually the weaker non-commutative little GT (see § 5) had amajor impact in Operator Algebra Cohomology (see [139]), starting with the proof of a conjectureof Ringrose in [112]. The non-commutative little GT is also connected and, in some sense equivalentto (see [94]), the non-commutative L1 Khintchine inequality, that we prove in § 9 using the recentpaper [48] that also contains best possible bounds on the constants. Let K ′

G (resp. k′G) denote

the best possible constant in the non-commutative GT (resp. the little GT). Curiously, in sharpcontrast with the commutative case, the exact values K ′

G = k′G = 2 are known, following [47]. We

present this in § 11.Although these results all deal with non-commutative C∗-algebras, they still belong to classical

Banach space theory. However, around 1988, a theory of non-commutative or “quantum” Banachspaces emerged with the thesis of Ruan and the work of Effros–Ruan, Blecher and Paulsen. Inthat theory the spaces remain Banach spaces but the morphisms are different: The familiar spaceB(E, F ) of bounded linear maps between two Banach spaces is replaced by the space CB(E, F )formed of the completely bounded (c.b. in short) ones. Moreover, each Banach space E comesequipped with an additional structure in the form of an isometric embedding (“realization”) E ⊂B(H) into the algebra of bounded operators on a Hilbert space H. Thus Banach spaces aregiven a structure resembling that of a C∗-algebra, but contrary to C∗-algebras which admit aprivileged realization, Banach spaces may admit many inequivalent operator space structures. Wegive a very brief outline of that theory in § 13. Through the combined efforts of Effros–Ruanand Blecher–Paulsen, an analogue of Grothendieck’s program was developed for operator spaces,including analogues of the injective and projective tensor products and the approximation property.In that process, a new kind of tensor product, the Haagerup one (denoted by ⊗h), gradually tookcenter stage. Although resembling Grothendieck’s Hilbertian tensor product ⊗H , it actually hasno true Banach space analogue. We describe it briefly in § 14. Later on, a bona fide analogue ofHilbert space (i.e. a unique self-dual object) denoted by OH was found (see §15) with factorizationproperties matching exactly those of classical Hilbert space (see [121]). Thus it became naturalto search for an analogue of GT for operator spaces. This came in several steps. First in [66], afactorization and extension theorem was proved for (jointly) c.b. bilinear forms on the product oftwo exact operator spaces, E, F ⊂ B(H), for instance two subspaces of the compact operators on H.This is described in § 16 and § 17. That result was surprising because it had no counterpart in theBanach space context, where nothing like that holds for subspaces of the space c0 of scalar sequencestending to zero. However, the main result was somewhat hybrid: it assumed complete boundednessbut only concluded to the existence of a bounded extension from E×F to B(H)×B(H). This wascorrected in [126]. There a characterization was found for jointly c.b. bilinear forms on E ×F withE, F exact. Curiously however, this factorization did not go through the canonical self-dual spaceOH —as one would have expected—but instead through a different Hilbertian space denoted byR ⊕ C, closely related to the Haagerup tensor product. In case E, F were C∗-algebras the resultestablished a conjecture formulated 10 years earlier by Effros–Ruan and Blecher. This however wasrestricted to exact C∗-algebras (or to suitably approximable bilinear forms). But in [49], Haagerupand Musat found a new approach that removed all restriction. Both [49, 126] have in common thatthey crucially use “Type III” von Neumann algebras. In § 18, we give an almost self-containedproof of the operator space GT, based on [49] but assuming no knowledge of Type III and hopefullymuch more accessible to a non-specialist. We also manage to incorporate in this approach the caseof c.b. bilinear forms on E×F with E, F exact operator spaces (from [126]), which was not coveredin [49].

2

In §19, we describe Tsirelson’s discovery of the close relationship between Grothendieck’s in-equality (i.e. GT) and Bell’s inequality. The latter was crucial to put to the test the Einstein–Podolsky–Rosen (EPR) theory of “hidden variables” proposed as a sort of substitute to quantummechanics. Using Bell’s ideas, experiments were made (see [6, 7]) to verify the presence of a certain“deviation” that invalidated the EPR theory. What Tsirelson observed is that the Grothendieckconstant can be interpreted as an upper bound for the “deviation” in the (generalized) Bell inequal-ities. This corresponds to an experiment with essentially two independent (because very distant)observers, and hence to the tensor product of two spaces. When three very far apart (and henceindependent) observers are present, the quantum setting leads to a triple tensor product, whencethe question whether there is a trilinear version of GT. We present the recent counterexamplefrom [34] to this “trilinear GT” in §20. We follow the same route as [34], but by using a differenttechnical ingredient, we are able to include a rather short self-contained proof.

Consider an n × n matrix [aij ] with real entries. Following Tsirelson [148], we say that [aij ] isa quantum correlation matrix if there are self-adjoint operators Ai, Bj on a Hilbert space H with‖Ai‖ ≤ 1, ‖Bj‖ ≤ 1 and ξ in the unit sphere of H ⊗2 H such that

(1.1) ∀i, j = 1, . . . , n aij = 〈(Ai ⊗ Bj)ξ, ξ〉.

If in addition the operators Ai | 1 ≤ i ≤ n and Bj | 1 ≤ j ≤ n all commute then [aij ] is calleda classical correlation matrix. In that case it is easy to see that there is a “classical” probabilityspace (Ω,A, P) and real valued random variables Ai, Bj in the unit ball of L∞ such that

(1.2) aij =

∫Ai(ω)Bj(ω) dP(ω).

As observed by Tsirelson, GT implies that any real matrix of the form (1.1) can be written in theform (1.2) after division by KR

G and this is the best possible constant (valid for all n). This isprecisely what (1.14) below says: Indeed, (1.1) (resp. (1.2)) holds iff the norm of

∑aijei ⊗ ej in

ℓn∞⊗H ℓn

∞ (resp. ℓn∞

∧⊗ ℓn

∞) is less than 1 (see the proof of Theorem 12.12 below for the identificationof (1.1) with the unit ball of ℓn

∞ ⊗H ℓn∞). In [148], Tsirelson, observing that in (1.1), Ai ⊗ 1 and

1 ⊗ Bj are commuting operators on H = H ⊗2 H, considered the following generalization of (1.1):

(1.3) ∀i, j = 1, . . . , n aij = 〈XiYjξ, ξ〉

where Xi, Yj ∈ B(H) with ‖Xi‖ ≤ 1 ‖Yj‖ ≤ 1 are self-adjoint operators such that XiYj = YjXi forall i, j and ξ is in the unit sphere of H. If we restrict to H and H finite dimensional, then (1.1) and(1.3) are equivalent, by fairly routine representation theory. Thus Tsirelson [146, Th. 1] initiallythought that (1.1) and (1.3) were actually the same in general, but he later on emphasized that hecompletely overlooked a serious approximation difficulty, and he advertised this, in a generalizedform, as problem 33 (see [149]) on a website (http://www.imaph.tu-bs.de/qi/problems/) devotedto quantum information theory, see [138] as a substitute for the website.

As it turns out (see [61, 32]) the generalized form of Tsirelson’s problem is equivalent to oneof the most famous ones in von Neumann algebra theory going back to Connes’s paper [22]. TheConnes problem can be stated as follows: Consider two unitaries U, V in a von Neumann algebraM equipped with a faithful normal trace τ with τ(1) = 1, can we find nets (Uα) (V α) of unitarymatrices of size N(α) × N(α) such that

τ(P (U, V )) = limα→∞

1

N(α)tr(P (Uα, V α))

3

for any polynomial P (X, Y ) (in noncommuting variables X, Y )?Note we can restrict to pairs of unitaries by a well known matrix trick (see [153, Cor. 2]).In [77], Kirchberg found many striking equivalent reformulations of this problem, among which thisone stands out: Is there a unique C∗-norm on the tensor product C ⊗ C when C is the (full) C∗-algebra of the free group FN with N ≥ 2 generators? The connection with GT comes through thegenerators: if U1, . . . , UN are the generators of FN viewed as sitting in C then E = span[U1, . . . , UN ]is C-isometric to (N -dimensional) ℓ1 and GT tells us that the minimal and maximal C∗-norms ofC ⊗ C are KC

G-equivalent on E ⊗ E.In Kirchberg’s result, one can replace C by the free product of n copies of C([−1, 1]), or equivalentlywe can assume that C is the universal C∗-algebra generated by an n-tuple x1, · · · , xn of self-adjointswith norm ≤ 1 (n ≥ 2). This can be connected to the difference between (1.1) and (1.3): Indeed,depending whether ϕ is a state on C⊗min C or on C⊗max C, the associated matrix aij = ϕ(xi⊗xj)is of the form (1.1) or of the form (1.3).In addition to this link with C∗-tensor products, the operator space version of GT has led to thefirst proof in [66] that B(H) ⊗ B(H) admits at least 2 inequivalent C∗-norms. We describe someof these results connecting GT to C∗-tensor products and the Connes–Kirchberg problem in §12.

Lastly, in §22 we briefly describe the recent surge of interest in GT among computer scientists,apparently triggered by the idea ([3]) to attach a Grothendieck inequality (and hence a Grothendieckconstant) to any (finite) graph. The original GT corresponds to the case of bipartite graphs. Themotivation for the extension lies in various algorithmic applications of the related computations.Here is a rough glimpse of the connection with GT: When dealing with certain “hard” optimiza-tion problems of a specific kind (“hard” here means time consuming), computer scientists have away to replace them by a companion problem that can be solved much faster using semidefiniteprogramming. The companion problem is then called the semidefinite “relaxation” of the originalone. For instance, consider a finite graph G = (V, E), and a real matrix [aij ] indexed by V × V .We propose to compute

(I) = max∑

i,j∈Eaijsisj | si = ±1, sj = ±1.

In general, computing such a maximum is hard, but the relaxed companion problem is to compute

(II) = max∑

i,j∈Eaij〈xi, yj〉 | xi ∈ BH , yj ∈ BH

where BH denotes the unit ball in Hilbert space H. The latter is much easier: It can be solved(up to an arbitrarily small additive error) in polynomial time by a well known method called the“ellipsoid method” (see [39]).The Grothendieck constant of the graph is defined as the best K such that (II) ≤ K(I). Of

course (I) ≤ (II). Thus the Grothendieck constant is precisely the maximum ratio relaxed(I)(I) . When

V is the disjoint union of two copies S′ and S′′ of [1, . . . , n] and E is the union of S′ × S′′ andS′′ × S′ (“bipartite” graph), then GT (in the real case) says precisely that (II) ≤ KG(I) (seeTheorem 2.4), so the constant KG is the maximum Grothendieck constant for all bipartite graphs.Curiously, the value of these constants can also be connected to the P=NP problem. We merelyglimpse into that aspect in §22, and refer the reader to the references for a more serious exploration.

The 3 main tensor norms Let X, Y be Banach spaces, let X ⊗ Y denote their algebraic tensorproduct. Then for any

(1.4) T =∑n

1xj ⊗ yj

4

in X ⊗ Y we define

(1.5) ‖T‖∧ = inf∑

‖xj‖‖yj‖

(“projective norm”)

or equivalently

(1.6) ‖T‖∧ = inf

(∑

‖xj‖2)1/2(∑

‖yj‖2)1/2

where the infimum runs over all possible representations of the form (1.4); furthermore

‖T‖∨ = sup

∣∣∣∑

x∗(xj)y∗(yj)

∣∣∣∣∣∣∣∣ x∗ ∈ BX∗ , y∗ ∈ BY ∗

(“injective norm”)(1.7)

‖T‖H = inf

sup

x∗∈BX∗

(∑|x∗(xj)|2

)1/2sup

y∗∈BY ∗

(∑|y∗(yj)|2

)1/2

(1.8)

where again the infimum runs over all possible representions (1.4).Note the obvious inequalities

(1.9) ‖T‖∨ ≤ ‖T‖H ≤ ‖T‖∧

the last one because of (1.6). Let T : X∗ → Y be the linear mapping associated to T , so thatT (x∗) =

∑x∗(xj)yj . Then

‖T‖∨ = ‖T‖B(X,Y )

and

(1.10) ‖T‖H = inf‖T1‖‖T2‖

where the infimum runs over all Hilbert spaces H and all possible factorizations of T through H:

T : X∗ T2−→ HT1−→ Y

with T = T1T2. When T : Z → Y is an operator between arbitrary Banach spaces, the right handside of (1.10) is usually denoted by γ2(T ) and called “the norm of factorization through Hilbertspace of T .”

One of the great methodological innovations of “the Resume” was the systematic use of dualityof tensor norms: Given a norm α on X ⊗ Y one defines α∗ on X∗ ⊗ Y ∗ by setting

∀T ′ ∈ X∗ ⊗ Y ∗ α∗(T ′) = sup|〈T, T ′〉| | T ∈ X ⊗ Y, α(T ) ≤ 1.

In the case α(T ) = ‖T‖H , Grothendieck studied the dual norm α∗ and used the notation α∗(T ) =‖T‖H′ . We have

‖T‖H′ = inf

(∑

‖xj‖2)1/2(∑

‖yj‖2)1/2

where the infimum runs over all finite sums∑n

1 xj ⊗ yj ∈ X ⊗ Y such that

∀(x∗, y∗) ∈ X∗ × Y ∗ |〈T, x∗ ⊗ y∗〉| ≤ (∑

|x∗(xj)|2)1/2(∑

|y∗(yj)|2)1/2.

It is easy to check that if α(T ) = ‖T‖∧ then α∗(T ′) = ‖T ′‖∨. Moreover, if either X or Y is finitedimensional and β(T ) = ‖T‖∨ then β∗(T ′) = ‖T ′‖∧. So, at least in the finite dimensional setting

5

the projective and injective norms ‖ ‖∧ and ‖ ‖∨ are in perfect duality (and so are ‖ ‖H and ‖ ‖H′).

The 3 fundamental Banach spaces The 3 spaces we have in mind are L2 (or any Hilbertspace), L∞ and L1. There is little doubt that Hilbert spaces are central, but Dvoretzky’s famoustheorem that any infinite dimensional Banach space contains almost isometric copies of any finitedimensional Hilbert space makes it all the more striking. As for L∞ and L1, their central role isattested by the fact that any separable Banach space embeds isometrically into L∞([0, 1]) or ℓ∞(resp. is isometric to a quotient of L1([0, 1]) or ℓ1). Of course, by L2, L∞ and L1 we think here ofinfinite dimensional spaces, but actually the discovery that the finite dimensional setting is crucialto the understanding of many features of the structure of infinite dimensional Banach spaces isanother visionary innovation of the resume (Grothendieck even conjectured explicitly Dvoretzky’s1961 theorem in the shorter article [37, p. 108] that follows the resume).Throughtout the paper (with K = R or C) ℓn

p designates Kn equipped with the norm ‖x‖ =

(Σ|xj |p)1/p and ‖x‖ = max |xj | when p = ∞. We denote by (e1, . . . , en) the canonical basis ofℓn∞ and by (e∗1, . . . , e

∗n) the biorthogonal basis in ℓn

1 = (ℓn∞)∗. Then, when S = [1, . . . , n], we have

C(S) = ℓn∞ and C(S)∗ = ℓn

1 . Any T ∈ C(S) ⊗ C(S) (resp. T ′ ∈ C(S)∗ ⊗ C(S)∗) can be identifiedwith a matrix [aij ] (resp. [a′ij ]) by setting

T = Σaijei ⊗ ej (resp. T ′ = Σa′ij ⊗ e∗i ⊗ e∗j ).

One then checks easily from the definitions that (recall K = R or C)

(1.11) ‖T ′‖∨ = sup

∣∣∣∑

a′ijαiβj

∣∣∣∣∣∣∣ αi, βj ∈ K, supi |αi| ≤ 1, supj |βj | ≤ 1

.

Moreover,

(1.12) ‖T‖H = infsupi

‖xi‖ supj

‖yj‖

where the infimum runs over all Hilbert spaces H and all xi, yj in H such that aij = 〈xi, yj〉 for alli, j = 1, . . . , n. By duality, this implies

(1.13) ‖T ′‖H′ = sup∣∣∣

∑a′ij〈xi, yj〉

∣∣∣

where the supremum is over all Hilbert spaces H and all xi, yj in the unit ball of H.

First statement of GT With these definitions, our fundamental result GT can be stated asfollows: there is a constant K such that for any T in L∞ ⊗L∞ (or any T in C(S)⊗C(S)) we have

(1.14) ‖T‖∧ ≤ K‖T‖H .

Equivalently by duality the theorem says that for any T ′ in L1 ⊗ L1 we have

(1.15) ‖T ′‖H′ ≤ K‖T ′‖∨.

The best constant in either (1.14) (or its equivalent dual form (1.15)) is denoted by KG andcalled the Grothendieck constant. Actually we must distinguish its value in the real case KR

G andin the complex case KC

G. To this day, its exact value is still unknown although it is known that1 < KC

G < KRG ≤ 1.782, see §3 for more information.

6

Focusing on the finite dimensional case, we find a more concrete but equivalent form of GT. Wemay as well restrict to the case when S is a finite set, say S = [1, . . . , n] so that C(S) = ℓn

∞ andC(S)∗ = ℓn

1 . Thus GT in the form of (1.15) can be rephrased as the domination of (1.13) by (1.11)multiplied by a fixed constant K independent of n. The best such K is equal to KG (see Theorem2.4 below).

Remark. By (1.9), (1.15) implies‖T‖H′ ≤ K‖T‖H .

but the latter inequality is much easier to prove than Grothendieck’s, so that it is often called “thelittle GT” and for it the best constant denoted by kG is known: it is equal to π/2 in the real caseand 4/π in the complex one, see §5.

2. Classical GT

In this section, we take the reader on a tour of the many reformulations of GT.

Theorem 2.1 (Classical GT/factorization). Let S, T be compact sets. For any bounded bilinearform ϕ : C(S) × C(T ) → K (here K = R or C) there are probabilities λ and µ, respectively on Sand T , such that

(2.1) ∀(x, y) ∈ C(S) × C(T ) |ϕ(x, y)| ≤ K‖ϕ‖(∫

S|x|2dλ

)1/2 (∫

T|y|2dµ

)1/2

where K is a numerical constant, the best value of which is denoted by KG, more precisely by KRG

or KCG depending whether K = R or C.

Equivalently, the linear map ϕ : C(S) → C(T )∗ associated to ϕ admits a factorization of the formϕ = J∗

µuJλ where Jλ : C(S) → L2(λ) and Jµ : C(T ) → L2(µ) are the canonical (norm 1) inclusionsand u : L2(λ) → L2(µ)∗ is a bounded linear operator with ‖u‖ ≤ K‖ϕ‖.

For any operator v : X → Y , we denote

(2.2) γ2(v) = inf‖v1‖‖v2‖

where the infimum runs over all Hilbert spaces H and all possible factorizations of v through H:

v : Xv2−→ H

v1−→ Y

with v = v1v2.Note that any L∞-space is isometric to C(S) for some S, and any L1-space embeds isometricallyinto its bidual, and hence embeds into a space of the form C(T )∗. Thus we may state:

Corollary 2.2. Any bounded linear map v : C(S) → C(T )∗ or any bounded linear map v : L∞ →L1 (over arbitrary measure spaces) factors through a Hilbert space. More precisely, we have

γ2(v) ≤ ℓ‖v‖

where ℓ is a numerical constant with ℓ ≤ KG.

By a Hahn–Banach type argument (see §4), the preceding theorem is equivalent to the followingone:

7

Theorem 2.3 (Classical GT/inequality). For any ϕ : C(S)×C(T ) → K and for any finite sequence(xj , yj) in C(S) × C(T ) we have

(2.3)∣∣∣∑

ϕ(xj , yj)∣∣∣ ≤ K‖ϕ‖

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

∥∥∥∥(∑

|yj |2)1/2

∥∥∥∥∞

.

(We denote ‖f‖∞ = supS

|f(s)| for f ∈ C(S).) Here again

Kbest = KG.

Assume S = T = [1, . . . , N ]. Note that C(S) = C(T ) = ℓN∞. Then we obtain the formulation

put forward by Lindenstrauss and PeÃlczynski in [89], perhaps the most “concrete” or elementaryof them all:

Theorem 2.4 (Classical GT/inequality/discrete case). Let [aij ] be an N×N scalar matrix (N ≥ 1)such that

(2.4) ∀α, β ∈ Kn

∣∣∣∑

aijαiβj

∣∣∣ ≤ supi

|αi| supj

|βj |.

Then for any Hilbert space H and any N -tuples (xi), (yj) in H we have

(2.5)∣∣∣∑

aij〈xi, yj〉∣∣∣ ≤ K sup ‖xi‖ sup ‖yj‖.

Moreover the best K (valid for all H and all N) is equal to KG.

Proof. We will prove that this is equivalent to the preceding Theorem 2.3. Let S = T = [1, . . . , N ].Let ϕ : C(S) × C(T ) → K be the bilinear form associated to [aij ]. Note that (by our assumption)‖ϕ‖ ≤ 1. We may clearly assume that dim(H) < ∞. Let (e1, . . . , ed) be an orthonormal basis ofH. Let

Xk(i) = 〈xi, ek〉 and Yk(j) = 〈yj , ek〉.Then

∑aij〈xi, yj〉 =

∑

k

ϕ(Xk, Yk),

sup ‖xi‖ =

∥∥∥∥(∑

|Xk|2)1/2

∥∥∥∥∞

and sup ‖yj‖ =

∥∥∥∥(∑

|Yk|2)1/2

∥∥∥∥∞

.

Then it is clear that Theorem 2.3 implies Theorem 2.4. The converse is also true. To see that oneshould view C(S) and C(T ) as L∞-spaces (with constant 1). This means that either X = C(S)or X = C(T ) (assume S, T compact) can be rewritten, for each fixed ε > 0, as X =

⋃α

Xα where

(Xα) is a net (directed by inclusion) of finite dimensional subspaces of X such that, for each α, Xα

is (1 + ε)-isomorphic to ℓN(α)∞ where N(α) = dim(Xα).

In harmonic analysis, the classical Marcinkiewicz–Zygmund inequality says that any boundedlinear map T : Lp(µ) → Lp(µ

′) satisfies the following inequality (here 0 < p ≤ ∞)

∀n ∀xj ∈ Lp(µ) (1 ≤ j ≤ n)

∥∥∥∥(∑

|Txj |2)1/2

∥∥∥∥p

≤ ‖T‖∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥p

.

8

Of course p = 2 is trivial. Moreover, the case p = ∞ (and hence p = 1 by duality) is obviousbecause we have the following linearization of the “square function norm”:

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

= sup∥∥∥

∑ajxj

∥∥∥∞

| aj ∈ K,∑

|aj |2 ≤ 1

.

The remaining case 1 < p < ∞ is an easy consequence of Fubini’s Theorem and the isometric linearembedding of ℓ2 into Lp provided by independent standard Gaussian variable (see (2.11) below):Indeed, if (gj) is an i.i.d. sequence of Gaussian normal variables relative to a probability P we havefor any sequence (xj) in Lp(µ)

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥p

= ‖g1‖−1p

(∫ ∥∥∥∑

gj(ω)xj

∥∥∥p

pdP(ω)

)1/p

.

The preceding result can be reformulated in a similar fashion:

Theorem 2.5 (Classical GT/Marcinkiewicz–Zygmund style). For any pair of measure spaces(Ω, µ), (Ω′, µ′) and any bounded linear map T : L∞(µ) → L1(µ

′) we have ∀n ∀xj ∈ L∞(µ)(1 ≤ j ≤ n)

(2.6)

∥∥∥∥(∑

|Txj |2)1/2

∥∥∥∥1

≤ K‖T‖∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

.

Moreover here again Kbest = KG.

Proof. The modern way to see that Theorems 2.3 and 2.5 are equivalent is to note that both resultscan be reduced by a routine technique to the finite case, i.e. the case S = T = [1, . . . , N ] = Ω = Ω′

with µ = µ′ = counting measure. Of course the constant K should not depend on N . In that casewe have C(S) = L∞(Ω) and L1(µ

′) = C(T )∗ isometrically, so that (2.3) and (2.6) are immediatelyseen to be identical using the isometric identity (for vector-valued functions) C(T ; ℓn

2 )∗ = L1(µ′; ℓn

2 ).Note that the reduction to the finite case owes a lot to the illuminating notion of L∞-spaces (see[89]).

Krivine [83] observed the following generalization of Theorem 2.5 (we state this as a corollary,but it is clearly equivalent to the theorem and hence to GT).

Corollary 2.6. For any pair Λ1, Λ2 of Banach lattices and for any bounded linear T : Λ1 → Λ2

we have

(2.7) ∀n ∀xj ∈ Λ1 (1 ≤ j ≤ n)

∥∥∥∥(∑

|Txj |2)1/2

∥∥∥∥Λ2

≤ K‖T‖∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥Λ1

.

Again the best K (valid for all pairs (Λ1, Λ2) is equal to KG.

Here the reader may assume that Λ1, Λ2 are “concrete” Banach lattices of functions over measurespaces say (Ω, µ), (Ω′, µ′) so that the notation ‖(∑ |xj |2)1/2‖Λ1 can be understood as the norm in Λ1

of the function ω → (∑ |xj(ω)|2)1/2. But actually this also makes sense in the setting of “abstract”

Banach lattices (see [83]).Grothendieck introduced the projective tensor product X⊗Y of two Banach spaces X, Y as the

completion of (X ⊗ Y, ‖ · ‖∧). Its characteristic property is the isometric identity

(X⊗Y )∗ = B(X × Y )

where B(X × Y ) denotes the space of bounded bilinear forms on X × Y .Then Theorem 2.3 admits the following dual reformulation:

9

Theorem 2.7 (Classical GT/predual formulation). For any F in C(S) ⊗ C(T ) we have

(2.8) ‖F‖∧ ≤ K‖F‖H

and, in C(S) ⊗ C(T ), (1.8) becomes

(2.9) ‖F‖H = inf

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

∥∥∥∥(∑

|yj |2)1/2

∥∥∥∥∞

with the infimum running over all n and all possible representations of F of the form

F =∑n

1xj ⊗ yj , (xj , yj) ∈ C(S) × S(T ).

We will now prove Theorem 2.7, and hence all the preceding equivalent formulations of GT.Note that both ‖ · ‖∧ and ‖ · ‖H are Banach algebra norms on C(S) ⊗ C(T ), with respect to thepointwise product on S × T , i.e. we have for any t1, t2 in C(S) ⊗ C(T )

(2.10) ‖t1 · t2‖∧ ≤ ‖t1‖∧‖t2‖∧ and ‖t1 · t2‖H ≤ ‖t1‖H‖t2‖H .

Let H = ℓ2. Let gj | j ∈ N be an i.i.d. sequence of standard Gaussian random variable on aprobability space (Ω,A, P). For any x =

∑xjej in ℓ2 we denote G(x) =

∑xjgj . Note that

(2.11) 〈x, y〉H = 〈G(x), G(y)〉L2(Ω,P).

Assume K = R. The following formula is crucial both to Grothendieck’s original proof and toKrivine’s: if ‖x‖H = ‖y‖H = 1

(2.12) 〈x, y〉 = sin(π

2〈sign(G(x)), sign(G(y))〉

).

Grothendieck’s proof of Theorem 2.7 with K = sh(π/2). This is in essence the original proof. Notethat Theorem 2.7 and Theorem 2.3 are obviously equivalent by duality. We already saw thatTheorem 2.3 can be reduced to Theorem 2.4 by an approximation argument (based on L∞-spaces).Thus it suffices to prove Theorem 2.7 in case S, T are finite subsets of the unit ball of H.

Then we may as well assume H = ℓ2. Let F ∈ C(S)⊗C(T ). We view F as a function on S×T .Assume ‖F‖H < 1. Then by definition of ‖F‖H , we can find elements xs, yt in ℓ2 with ‖xs‖ ≤ 1,‖yt‖ ≤ 1 such that

∀(s, t) ∈ S × T F (s, t) = 〈xs, yt〉.By adding mutually orthogonal parts to the vectors xs and yt, F (s, t) does not change and wemay assume ‖xs‖ = 1, ‖yt‖ = 1. By (2.12) F (s, t) = sin(π

2

∫ξsηt dP) where ξs = sign(G(xs)) and

ηt = sign(G(yt)).Let k(s, t) =

∫ξsηt dP. Clearly ‖k‖C(S)⊗C(T ) ≤ 1 this follows by approximating the integral by

sums (note that, S, T being finite, all norms are equivalent on C(S) ⊗ C(T ) so this approximationis easy to check). Since ‖ ‖∧ is a Banach algebra norm (see (2.10)) , the elementary identity

sin(z) =∑∞

0(−1)m z2m+1

(2m + 1)!

is valid for all z in C(S)⊗C(T ), and we have

‖ sin(z)‖∧ ≤∑∞

0‖z‖2m+1((2m + 1)!)−1 = sh(‖z‖∧).

Applying this to z = (π/2)k, we obtain

‖F‖∧ ≤ sh(π/2).

10

Krivine’s proof of Theorem 2.7 with K = π(2 Log(1 +√

2))−1. Let K = π/2a where a > 0 is cho-sen so that sh(a) = 1, i.e. a = Log(1 +

√2). We will prove Theorem 2.7 (predual form of Theorem

2.3). From what precedes, it suffices to prove this when S, T are arbitrary finite sets.Let F ∈ C(S) ⊗ C(T ). We view F as a function on S × T . Assume ‖F‖H < 1. We will prove

that ‖F‖∧ ≤ K. Since ‖ ‖H also is a Banach algebra norm (see (2.10)) we have

‖ sin(aF )‖H ≤ sh(a‖F‖H) < sh(a).

By definition of ‖ ‖H (after a slight correction, as before, to normalize the vectors), this meansthat there are (xs)s∈S and (yt)t∈T in the unit sphere of H such that

sin(aF (s, t)) = 〈xs, yt〉.

By (2.12) we have

sin(aF (s, t)) = sin

(π

2

∫ξsηt dP

)

where ξs = sign(G(xs)) and ηt = sign(G(yt)). Observe that |F (s, t)| ≤ ‖F‖H < 1, and a fortiori,since a < 1, |aF (s, t)| < π/2. Therefore we must have

aF (s, t) =π

2

∫ξsηt dP

and hence ‖aF‖∧ ≤ π/2, so that we conclude ‖F‖∧ ≤ π/2a.

We now give the Schur multiplier formulation of GT. By a Schur multiplier, we mean here abounded linear map T : B(ℓ2) → B(ℓ2) of the form T ([aij ]) = [ϕijaij ] for some (infinite) matrixϕ = [ϕij ]. We then denote T = Mϕ. For example, if ϕij = z′iz

′′j with |z′i| ≤ 1, |z′′j | ≤ 1 then

‖Mϕ‖ ≤ 1 (because Mϕ(a) = D′aD′′ where D′ and D′′ are the diagonal matrices with entries (z′i)and (z′′j )). Let us denote by S the class of all such “simple” multipliers. Let conv(S) denote thepointwise closure (on N × N) of the convex hull of S. Clearly ϕ ∈ S ⇒ ‖Mϕ‖ ≤ 1. Surprisingly,the converse is essentially true (up to a constant). This is one more form of GT:

Theorem 2.8 (Classical GT/Schur multipliers). If ‖Mϕ‖ ≤ 1 then ϕ ∈ KG conv(S) and KG is thebest constant satisfying this (by KG we mean KR

G or KCG depending whether all matrices involved

have real or complex entries).

We will use the following (note that, by (1.12), this implies in particular that ‖ ‖H is a Banachalgebra norm on ℓn

∞ ⊗ ℓn∞ since we obviously have ‖Mϕψ‖ = ‖MϕMψ‖ ≤ ‖Mϕ‖‖Mψ‖):

Proposition 2.9. We have ‖Mϕ‖ ≤ 1 iff there are xi, yj in the unit ball of Hilbert space such thatϕij = 〈xi, yj〉.

Remark. Except for precise constants, both Proposition 2.9 and Theorem 2.8 are due to Grothendieck,but they were rediscovered several times, notably by John Gilbert and Uffe Haagerup. They canbe essentially deduced from [36, Prop. 7, p. 68] in the Resume, but the specific way in whichGrothendieck uses duality there introduces some extra numerical factors (equal to 2) in the con-stants involved, that were removed later on (in [86]).

Proof of Theorem 2.8. Taking the preceding Proposition for granted, it is easy to complete theproof. Assume ‖Mϕ‖ ≤ 1. Let ϕn be the matrix equal to ϕ in the upper n × n corner andto 0 elsehere. Then ‖Mϕn‖ ≤ 1 and obviously ϕn → ϕ pointwise. Thus, it suffices to show

11

ϕ ∈ KG conv(S) when ϕ is an n × n matrix. Then let T =∑n

i,j=1 ϕijei ⊗ ej ∈ ℓn∞ ⊗ ℓn

∞. If‖Mϕ‖ ≤ 1, the preceding Proposition implies by (1.12) that ‖T‖H ≤ 1 and hence by (1.14) that‖T‖∧ ≤ KG. But by (1.5), ‖T‖∧ ≤ KG iff ϕ ∈ KG conv(S). A close examination of the proofshows that the best constant in Theorem 2.8 is equal to the best one in (1.14).

Proof of Proposition 2.9. The if part is easy to check. To prove the converse, we may again assume(e.g. using ultraproducts) that ϕ is an n × n matrix. Assume ‖Mϕ‖ ≤ 1. By duality it suffices toshow that |∑ ϕijψij | ≤ 1 whenever ‖∑

ψijei ⊗ ej‖ℓn1⊗H′ℓn

1≤ 1. Let T ′ =

∑ψijei ⊗ ej ∈ ℓn

1 ⊗ ℓn1 .

We will use the fact that ‖T ′‖H′ ≤ 1 iff ψ admits a factorization of the form ψij = αiaijβj with[aij ] in the unit ball of B(ℓn

2 ) and (αi), (βj) in the unit ball of ℓn2 . Using this fact we obtain

∣∣∣∑

ψijϕij

∣∣∣ =∣∣∣∑

λiaijϕijµj

∣∣∣ ≤ ‖[aijϕij ]‖ ≤ ‖Mϕ‖‖[aij ]‖ ≤ 1.

The preceding factorization of ψ requires a Hahn–Banach argument that we provide in Remark 4.4below.

Theorem 2.10. The constant KG is the best constant K such that, for any Hilbert space H, thereis a probability space (Ω, P) and functions

Φ: H → L∞(Ω, P), Ψ: H → L∞(Ω, P)

such that

∀x ∈ H ‖Φ(x)‖∞ ≤ ‖x‖, ‖Ψ(x)‖∞ ≤ ‖x‖

and

(2.13) ∀x, y ∈ H1

K〈x, y〉 =

∫Φ(x)Ψ(y) dP

Note that depending whether K = R or C we use real or complex valued L∞(Ω, P). Actually, wecan find a compact set Ω equipped with a Radon probability measure and functions Φ, Ψ as abovebut taking values in the space C(Ω) of continuous (real or complex) functions on Ω.

Proof. We will just prove that this holds for the constant K appearing in Theorem 2.7. It sufficesto prove the existence of functions Φ1, Ψ1 defined on the unit sphere of H satisfying the requiredproperties. Indeed, we may then set Φ(0) = Ψ(0) = 0 and

Φ(x) = ‖x‖Φ1(x‖x‖−1), Ψ(y) = ‖y‖Ψ1(y‖y‖−1)

and we obtain the desired properties on H. Our main ingredient is this: let S denote the unitsphere of H, let S ⊂ S be a finite subset, and let

∀(x, y) ∈ S × S FS(x, y) = 〈x, y〉.

Then FS is a function on S × S but we may view it as an element of C(S)⊗C(S). We will obtain(2.13) for all x, y in S and then use a limiting argument.

Obviously ‖FS‖H ≤ 1. Indeed let (e1, . . . , en) be an orthonormal basis of the span of S. Wehave

FS(x, y) =∑n

1xj yj

12

and supx∈S(∑ |xj |2)1/2 = supy∈S(

∑ |yj |2)1/2 = 1 (since (∑ |xj |2)1/2 = ‖x‖ = 1 for all x in S).

Therefore, ‖F‖H ≤ 1 and Theorem 2.7 implies

‖FS‖∧ ≤ K.

Let C denote the unit ball of ℓ∞(S) equipped with the weak-∗ topology. Note that C is compactand for any x in S, the mapping f 7→ f(x) is continuous on C. We claim that for any ε > 0 andany finite subset S ⊂ S there is a probability λ on C × C (depending on (ε, S)) such that

(2.14) ∀x, y ∈ S1

K(1 + ε)〈x, y〉 =

∫

C×C

f(x)g(y) dλ(f, g).

Indeed, since ‖FS‖∧ < K(1 + ε) by homogeneity we can rewrite 1K(1+ε)FS as a finite sum

1

K(1 + ε)FS =

∑λm fm ⊗ gm

where λm ≥ 0,∑

λm = 1, ‖fm‖C(S) ≤ 1, ‖gm‖C(S) ≤ 1. Let fm ∈ C, gm ∈ C denote the extensionsof fm and gm vanishing outside S. Setting λ =

∑λmδ

( efm,egm), we obtain the announced claim. We

view λ = λ(ε, S) as a net, where we let ε → 0 and S ↑ S. Passing to a subnet, we may assume thatλ = λ(ε, S) converges weakly to a probability P on C ×C. Let Ω = C ×C. Passing to the limit in(2.14) we obtain

∀x, y ∈ S 1

K〈x, y〉 =

∫

Ω

f(x)g(y) dP(f, g).

Thus if we set ∀ω = (f, g) ∈ Ω

Φ(x)(ω) = f(x) and Ψ(y)(ω) = g(y)

we obtain finally (2.13).To show that the best K in Theorem 2.10 is equal to KG, it suffices to show that Theorem 2.10

implies Theorem 2.4 with the same constant K. Let (aij), (xi), (yj) be as in Theorem 2.4. We have

〈xi, yj〉 = K

∫Φ(xi)Ψ(yj) dP

and hence

∣∣∣∑

aij〈xi, yj〉∣∣∣ = K

∣∣∣∣∫ (∑

aijΦ(xi)Ψ(yj))

dP

∣∣∣∣

≤ K sup ‖xi‖ sup ‖yj‖

where at the last step we used the assumption on (aij) and |Φ(xi)(ω)| ≤ ‖xi‖, |Ψ(yj)(ω)| ≤ ‖yj‖ foralmost all ω. Since KG is the best K appearing in either Theorem 2.7 or Theorem 2.4, this completesthe proof, except for the last assertion, but that follows from the isometry L∞(Ω, P) ≃ C(S) forsome compact set S (namely the spectrum of L∞(Ω, P)) and that allows to replace the integralwith respect to P by a Radon measure on S.

13

Remark 2.11. The functions Φ, Ψ appearing above are highly non-linear. In sharp contrast, wehave

∀x, y ∈ ℓ2 〈x, y〉 =

∫G(x)G(y) dP

and for any p < ∞‖G(x)‖p = ‖x‖ γ(p)

where γ(p) = (E|g1|p)1/p. Here x 7→ G(x) is linear but since γ(p) → ∞ when p → ∞, this formuladoes not produce a uniform bound for the norm in C(S)⊗C(S) of FS(x, y) = 〈x, y〉 with S ⊂ Sfinite.

Remark 2.12. It is natural to wonder whether one can take Φ = Ψ in Theorem 2.10 (possibly witha worse K). Surprisingly the answer is negative, [72]. More precisely, Kashin and Szarek estimatedthe best constant K(n) with the following property: for any x1, . . . , xn in the unit ball of H thereare functions Φi in the unit ball of L∞(Ω, P) on a probability space (Ω, P) (we can easily reduceconsideration to the Lebesgue interval) such that

∀i, j = 1, . . . , n1

K(n)〈xi, xj〉 =

∫ΦiΦjdP.

They showed that K(n) ≥ δ(Log n)1/2, but the exact order of growth was obtained in [3]:

K(n) ≥ δ(log(n)

for some δ > 0 independent of n. Note that since E supi≤n

|G(xi)| . c(Log n)1/2, it is rather easy to

see that K(n) ≤ c′logn. The fact that K(n) → ∞ answered a question raised by Megretski (see[99]) in connection with possible electrical engineering applications.

Remark 2.13. Similar questions were also considered long before in Menchoff’s work on orthogonalseries of functions. In that direction, if we assume, in addition to ‖xj‖H ≤ 1, that

∀(αj) ∈ Rn

∥∥∥∑

αjxj

∥∥∥H

≤(∑

|αj |2)1/2

then it is unknown whether this modified version of K(n) remains bounded when n → ∞. Thisproblem (due to Olevskii, see [103]) is a well known open question in the theory of boundedorthogonal systems (in connection with a.s. convergence of the associated series).

Among the many applications of GT, the following isomorphic characterization of Hilbert spacesis rather puzzling in view of the many related questions (mentioned below) that remain open tothis day.

It will be convenient to use the Banach–Mazur “distance” between two Banach spaces B1, B2,defined as follows:

d(B1, B2) = inf‖u‖ ‖u−1‖where the infimum runs over all isomorphisms u : B1 → B2, and we set d(B1, B2) = +∞ if thereis no such isomorphism.

Corollary 2.14. The following properties of a Banach space B are equivalent:

(i) Both B and its dual B∗ embed isomorphically into an L1-space.

(ii) B is isomorphic to a Hilbert space.

14

More precisely, if X ⊂ L1, and Y ⊂ L1 are L1-subspaces, with B ≃ X and B∗ ≃ Y ; then there isa Hilbert space H such that

d(B, H) ≤ KGd(B, X)d(B∗, Y ).

Here KG is KRG or KC

G depending whether K = R or C.

Proof. Using the Gaussian isometric embedding x 7→ G(x) it is immediate that any Hilbert spacesay H = ℓ2(I) embeds isometrically into L1 and hence, since H ≃ H∗ isometrically, (ii) ⇒ (i) isimmediate. Conversely, assume (i). Let v : B → L1 and w : B∗ → L1 denote the isomorphicembeddings. We may apply GT to the composition

u = wv∗ : L∞v∗

−→ B∗ w−→ L1.

By Corollary 2.2, wv∗ factors through a Hilbert space. Consequently, since w is an embedding, v∗

itself must factor through a Hilbert space and hence, since v∗ is onto, B∗ (a fortiori B) must beisomorphic to a Hilbert space. The last assertion is then easy to check.

Note that in (ii) ⇒ (i), even if we assume B, B∗ both isometric to subspaces of L1, we onlyconclude that B is isomorphic to Hilbert space. The question was raised already by Grothendieckhimself in [36, p. 66] whether one could actually conclude that B is isometric to a Hilbert space.Bolker also asked the same question in terms of zonoids (a symmetric convex body is a zonoid iffthe normed space admitting the polar body for its unit ball embeds isometrically in L1.) This wasanswered negatively by Rolf Schneider [137] but only in the real case: He produced n-dimensionalcounterexamples over R for any n ≥ 3. Rephrased in geometric language, there are (symmetric)zonoids in R

n whose polar is a zonoid but that are not ellipsoids. Note that in the real case thedimension 2 is exceptional because any 2-dimensional space embeds isometrically into L1 so thereare obvious 2D-counterexamples (e.g. 2-dimensional ℓ1). But apparently (as pointed out by J.Lindenstrauss) the infinite dimensional case, both for K = R or C is open, and also the complexcase seems open in all dimensions.

3. The Grothendieck constants

The constant KG is “the Grothendieck constant.” Its exact value is still unknown!Grothendieck [36] proved that π/2 ≤ KR

G ≤ sh(π/2) = 2.301... Actually, his argument (see (5.2)below for a proof) shows that ‖g‖−2

1 ≤ KG where g is a standard Gaussian variable such that Eg = 0and E|g|2 = 1. In the real case ‖g‖1 = E|g| = (2/π)1/2. In the complex case ‖g‖1 = (π/4)1/2, andhence KC

G ≥ 4/π. It is known (cf. [112]) that KCG < KR

G. Krivine ([84]) proved that

(3.1) KRG ≤ π/(2 Log(1 +

√2)) = 1.782 . . .

and conjectured that this is the exact value. The latter conjecture remains open, although heproved that KR

G ≥ 1.66 (unpublished, see also [133]).See the forthcoming paper [96] for a (possibly computer assisted) attempt to show that Krivine’s

bound is not optimal.In the complex case the corresponding bounds are due to Haagerup and Davie [23, 46]. Curi-

ously, there are difficulties to “fully” extend Krivine’s proof to the complex case. Following thispath, Haagerup [46] finds

(3.2) KCG ≤ 8

π(K0 + 1)= 1.4049...

15

where K0 is is the unique solution in (0, 1) of the equation

K

∫ π/2

0

cos2 t

(1 − K2 sin2 t)1/2dt =

π

8(K + 1).

Davie (unpublished) improved the lower bound to KCG > 1.338.

Nevertheless, Haagerup [46] conjectures that a “full” analogue of Krivine’s argument shouldyield a slightly better upper bound, he conjectures that:

KCG ≤

(∫ π/2

0

cos2 t

(1 + sin2 t)1/2dt

)−1

= 1.4046...!

Define ϕ : [−1, 1] → [−1, 1] by

ϕ(s) = s

∫ π/2

0

cos2 t

(1 − s2 sin2 t)1/2dt.

Then (see [46]), ϕ−1 : [−1, 1] → [−1, 1] exists and admits a convergent power series expansion

ϕ−1(s) =∞∑

k=0

β2k+1s2k+1,

with β1 = 4/π and β2k+1 ≤ 0 for k ≥ 1. On one hand K0 is the solution of the equation2β1ϕ(s) = 1 + s, or equivalently ϕ(s) = π

8 (s + 1), but on the other hand Haagerup’s alreadymentioned conjecture in [46] is that, extending ϕ analytically, we should have

KCG ≤ |ϕ(i)| =

(∫ π/2

0

cos2 t

(1 + sin2 t)1/2dt

)−1

.

Konig [81] pursues this and obtains several bounds for the finite dimensional analogues of KCG,

which we now define.Let KR

G(N, d) and KCG(N, d) be the best constant K such that (2.5) holds (for all matrices [aij ]

satisfying (2.4)) for all N -tuples (x1, . . . , xN ), (y1, . . . , yN ) in a d-dimensional Hilbert space H onK = R or C respectively. Note that KK

G(N, d) ≤ KKG(N, N) for any d (since we can replace (xj),

(yj) by (Pxj) , (Pyj), P being the orthogonal projection onto span[xj ]).We set

(3.3) KKG(d) = sup

N≥1KK

G(N, d).

Equivalently, KKG(d) is the best constant in (2.5) when one restricts the left hand side to d-

dimensional unit vectors (xi, yj) (and N is arbitrary).Krivine [85] proves that KR

G(2) =√

2, KRG(4) ≤ π/2 and obtains numerical upper bounds for

KRG(N) for each N . The fact that KR

G(2) ≥√

2 is obvious since, in the real case, the 2-dimensionalL1- and L∞-spaces (namely ℓ2

1 and ℓ2∞) are isometric, but at Banach-Mazur distance

√2 from ℓ2

2.The assertion KR

G(2) ≤√

2 equivalently means that for any R-linear T : L∞(µ; R) → L1(µ′; R) the

complexification T C satisfies

‖T C : L∞(µ; C) → L1(µ′; C)‖ ≤

√2‖T : L∞(µ; R) → L1(µ

′; R)‖.

16

In the complex case, refining Davie’s lower bound, Konig [81] obtains two sided bounds (interms of the function ϕ) for KC

G(d), for example he proves that 1.1526 < KCG(2) < 1.2157. He also

computes the d-dimensional version of the preceding Haagerup conjecture on KCG. See [81, 82] for

more details, and [31] for some explicitly computed examples. See also [23] for various remarkson the norm ‖[aij ]‖(p) defined as the supremum of the left hand side of (2.5) over all unit vectorsxi, yj in an p-dimensional complex Hilbert space. In particular, Davie, Haagerup and Tonge (see[23, 144]) independently proved that (2.5) restricted to 2 × 2 complex matrices holds with K = 1.

Recently, in [17], the following variant of KRG(d) was introduced and studied: let KG[d, m]

denote the smallest constant K such that for any real matrix [aij ] of arbitrary size

sup |∑

aij〈xi, yj〉| ≤ K sup |∑

aij〈x′i, y

′j〉|

where the supremum on the left (resp. right) runs over all unit vectors (xi, yj) (resp. (x′i, y

′j) ) in

an d-dimensional (resp. m-dimensional) Hilbert space. Clearly we have KRG(d) = KG[d, 1].

The best value ℓbest of the constant in Corollary 2.2 seems unknown in both the real and complexcase. Note that, in the real case, we have obviously ℓbest ≥

√2 because, as we just mentioned, the

2-dimensional L1 and L∞ are isometric.

4. The Hahn-Banach argument

Grothendieck used duality of tensor norms and doing that, of course he used the Hahn-Banachextension theorem repeatedly. However, he did not use it in the way that is described below, e.g.to describe the H ′-norm. At least in [36], he systematically passes to the “self-adjoint” and “positivedefinite” setting and uses positive linear forms. This explains why he obtains only ‖.‖H ≤ 2‖.‖H′ ,and, although he suspects that it is true without it, he cannot remove the factor 2. This was donelater on, in Pietsch’s ([111]) and Kwapien’s work (see in particular [86]). Pietsch’s factorizationtheorem for p-absolutely summing operators became widely known. That was proved using a formof the Hahn-Banach separation, that has become routine to Banach space specialists since then.Since this kind of argument is used repeatedly in the sequel, we devote this section to it, but itwould fit better as an appendix.

We start with a variant of the famous min-max lemma.

Lemma 4.1. Let S be a set and let F ⊂ ℓ∞(S) be a convex cone of real valued functions on S suchthat

∀ f ∈ F sups∈S

f(s) ≥ 0.

Then there is a net (λα) of finitely supported probability measures on S such that

∀ f ∈ F lim

∫fdλα ≥ 0.

Proof. Let ℓ∞(S, R) denote the space all bounded real valued functions on S with its usual norm.In ℓ∞(S, R) the set F is disjoint from the set C− = ϕ ∈ ℓ∞(S, R) | sup ϕ < 0. Hence bythe Hahn-Banach theorem (we separate the convex set F and the convex open set C−) there isa non zero ξ ∈ ℓ∞(S, R)∗ such that ξ(f) ≥ 0 ∀ f ∈ F and ξ(f) ≤ 0 ∀ f ∈ C−. Let M ⊂ℓ∞(S, R)∗ be the cone of all finitely supported (nonnegative) measures on S viewed as functionalson ℓ∞(S, R). Since we have ξ(f) ≤ 0 ∀ f ∈ C−, ξ must be in the bipolar of M for the duality ofthe pair (ℓ∞(S, R), ℓ∞(S, R)∗). Therefore, by the bipolar theorem, ξ is the limit for the topologyσ(ℓ∞(S, R)∗, ℓ∞(S, R)) of a net of finitely supported (nonnegative) measures ξα on S. We have for

17

any f in ℓ∞(S, R), ξα(f) → ξ(f) and this holds in particular if f = 1, thus (since ξ is nonzero) wemay assume ξα(1) > 0, hence if we set λα(f) = ξα(f)/ξα(1) we obtain the announced result.

The next statement is meant to illustrate the way the preceding Lemma is used, but manyvariants are possible. We hope the main idea is sufficiently clear for later reference.

Note that if B1, B2 below are unital and commutative, we can identify them with C(T1), C(T2)for compact sets T1, T2. A state on B1 (resp. B2) then corresponds to a (Radon) probabilitymeasure on T1 (resp. T2).

Proposition 4.2. Let B1, B2 be C∗-algebras, let F1 ⊂ B1 and F2 ⊂ B2 be two linear subspaces,and let ϕ : F1 × F2 → C be a bilinear form. The following are equivalent:

(i) For any finite sets (xj1) and (xj

2) in F1 and F2 respectively, we have

∣∣∣∑

ϕ(xj1, x

j2)

∣∣∣ ≤∥∥∥∑

xj1x

j∗1

∥∥∥1/2 ∥∥∥

∑xj∗

2 xj2

∥∥∥1/2

.

(ii) There are states f1 and f2 on B1 and B2 respectively such that

∀(x1, x2) ∈ F1 × F2 |ϕ(x1, x2)| ≤ (f1(x1x∗1)f2(x

∗2x2))

1/2.

(iii) The form ϕ extends to a bounded bilinear form ϕ satisfying (i),(ii) on B1 × B2.

Proof. Assume (i). First observe that by the arithmetic/geometric mean inequality we have forany a, b ≥ 0

(ab)1/2 = inft>0

2−1(ta + (b/t)).

In particular we have

∥∥∥∑

xj1x

j∗1

∥∥∥1/2 ∥∥∥

∑xj∗

2 xj2

∥∥∥1/2

≤ 2−1(∥∥∥

∑xj

1xj∗1

∥∥∥ +∥∥∥∑

xj∗2 xj

2

∥∥∥)

.

Let Si be the set of states on Bi (i = 1, 2) and let S = S1 × S2. The last inequality implies

∣∣∣∑

ϕ(xj1, x

j2)

∣∣∣ ≤ 2−1 supf=(f1,f2)∈S

f1

(∑xj

1xj∗1

)+ f2

(∑xj∗

2 xj2

).

Moreover, since the right side does not change if we replace x1j by zjx

1j with zj ∈ C arbitrary

such that |zj | = 1, we may assume that the last inequality holds with∑ |ϕ(xj

1, xj2)| instead of∣∣∣

∑ϕ(xj

1, xj2)

∣∣∣. Then let F ⊂ ℓ∞(S, R) be the convex cone formed of all possible functions F : S → R

of the formF (f1, f2) =

∑

j

2−1f1(xj1x

j∗1 ) + 2−1f2(x

j∗2 xj

2) − |ϕ(xj1, x

j2)|.

By the preceding Lemma, there is a net U of probability measures (λα) on S such that for any Fin F we have

limU

∫F (g1, g2)dλα(g1, g2) ≥ 0.

We may as well assume that U is an ultrafilter. Then if we set

fi = limU

∫gidλα(g1, g2) ∈ Si

18

(in the weak-∗ topology σ(B∗i , Bi)), we find that for any choice of (xj

1) and (xj2) we have

∑

j

2−1f1(xj1x

j∗1 ) + 2−1f2(x

j∗2 xj

2) − |ϕ(xj1, x

j2)| ≥ 0

hence in particular ∀x1 ∈ F1,∀x2 ∈ F2

2−1(f1(x1x∗1) + f2(x

∗2x2)) ≥ |ϕ(x1, x2)|.

By the homogeneity of ϕ, this implies

inft>0

2−1(tf1(x1x∗1) + f2(x

∗2x2)/t) ≥ |ϕ(x1, x2)|

hence we obtain the desired conclusion (ii) using our initial observation on the geometric/arithmeticmean inequality. This shows (i) implies (ii). The converse is obvious. To show that (ii) implies(iii), let H1 (resp. H2) be the Hilbert space obtained from equipping B1 (resp. B2) with the scalarproduct 〈x, y〉 = f1(y

∗x) (resp. = f2(y∗x)) (“GNS construction”), let Jk : Bk → Hk (k = 1, 2) be

the canonical inclusion, and let Sk = Jk(Ek) ⊂ Hk. By (ii) ϕ defines a bilinear form of norm ≤ 1on S1 × S2. Using the orthogonal projection from say H1 to S1, the latter extends to a form ψ ofnorm ≤ 1 on H1 × H2, and then ϕ(x1, x2) = ψ(J1(x1), J2(x2)) is the desired extension.

Let Tj denote the unit ball of F ∗j equipped with the (compact) weak-∗ topology. Applying the

preceding to the embedding Fj ⊂ Bj = C(Tj) and recalling (1.8), we find

Corollary 4.3. Let ϕ ∈ (F1 ⊗ F2)∗. The following are equivalent:

(i) We have |ϕ(t)| ≤ ‖t‖H (∀t ∈ F1 ⊗ F2).

(ii) There are probabilities λ1 and λ2 on T1 and T2 respectively such that

∀(x1, x2) ∈ F1 × F2 |ϕ(x1, x2)| ≤ (

∫|x1|2dλ1

∫|x2|2dλ2)

1/2.

(iii) The form ϕ extends to a bounded bilinear form ϕ satisfying (i),(ii) on B1 ⊗ B2.

Remark 4.4. Consider a bilinear form ψ on ℓn∞ ⊗ ℓn

∞ defined by ψ(ei ⊗ ej) = ψij . The precedingCorollary implies that ‖ψ‖H′ ≤ 1 iff there are (αi), (βj) in the unit sphere of ℓn

2 such that for anyx1, x2 in ℓn

2 ∣∣∣∑

ψijx1(i)x2(j)∣∣∣ ≤

(∑|αi|2|x1(i)|2

)1/2 (∑|βj |2|x2(j)|2

)1/2.

Thus ‖ψ‖H′ ≤ 1 iff we can write ψij = αiaijβj for some matrix [aij ] of norm ≤ 1 on ℓn2 , and some

(αi), (βj) in the unit sphere of ℓn2 .

Here is another variant:

Proposition 4.5. Let B be a C∗-algebra, let F ⊂ B be a linear subspace, and let u : F → E be alinear map into a Banach space E. Fix numbers a, b ≥ 0. The following are equivalent:

(i) For any finite sets (xj) in F , we have

(∑‖uxj‖2

)1/2≤

(a

∥∥∥∑

xjx∗j

∥∥∥ + b∥∥∥∑

x∗jxj

∥∥∥)1/2

.

19

(ii) There are states f, g on B such that

∀x ∈ F ‖ux‖ ≤ (af(xx∗) + bg(x∗x))1/2.

(iii) The map u extends to a linear map u : B → E satisfying (i) (or (ii)) on the whole of B.

Proof. Let S be the set of pairs (f, g) of states on B. Apply Lemma 4.1 to the family of functionson S of the form (f, g) 7→ a

∑f(xjx

∗j )+ b

∑f(x∗

jxj)−∑ ‖uxj‖2. The extension in (iii) is obtained

using the orthogonal projection onto the space spanned by F in the Hilbert space associated to theright hand side of (ii). We leave the easy details to the reader.

Remark 4.6. When B is commutative, i.e. B = C(T ) with T compact, the inequality in (i) becomes,letting c = (a + b)1/2

(4.1)(∑

‖uxj‖2)1/2

≤ c∥∥∥∑

|xj |2∥∥∥

1/2.

Using F ⊂ C(T ) when T is equal to the unit ball of F ∗, we find the “Pietsch factorization” of u:there is a probability λ on T such that ‖u(x)‖2 ≤ c2

∫|x|2dλ ∀x ∈ C(T ). An operator satisfying

this for some c is called 2-summing and the smallest constant c for which this holds is denoted byπ2(u). More generally, if 0 < p < ∞, u is called p-summing with πp(u) ≤ c if there is a probabilityλ on T such that ‖u(x)‖p ≤ cp

∫|x|pdλ ∀x ∈ C(T ), but we will really use only p = 2 in this paper.

See [107] for a recent use of this Pietsch factorization for representations of H∞.

5. The “little” GT

The next result, being much easier to prove that GT has been nicknamed the “little GT.”

Theorem 5.1. Let S, T be compact sets. There is an absolute constant k such that, for any pairof bounded linear operators u : C(S) → H, v : C(T ) → H into an arbitrary Hilbert space H andfor any finite sequence (xj , yj) in C(S) × C(T ) we have

(5.1)∣∣∣∑

〈u(xj), v(yj)〉∣∣∣ ≤ k‖u‖‖v‖

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

∥∥∥∥(∑

|yj |2)1/2

∥∥∥∥∞

.

Let kG denote the best possible such k. We have kG = ‖g‖−21 (recall g denotes a standard N(0, 1)

Gaussian variable) or more explicitly

(5.2) kRG = π/2 kC

G = 4/π.

Although he used a different formulation and denoted kRG by σ, Grothendieck did prove that

kRG = π/2 (see [36, p.51]). His proof extends immediately to the complex case.

To see that Theorem 5.1 is “weaker” than GT (in the formulation of Theorem 2.3), let ϕ(x, y) =〈u(x), v(y)〉. Then ‖ϕ‖ ≤ ‖u‖‖v‖ and hence (2.3) implies Theorem 5.1 with

(5.3) kG ≤ KG.

Here is a very useful reformulation of the little GT:

Theorem 5.2. Let S be a compact set. For any bounded linear operator u : C(S) → H, thefollowing holds:

20

(i) There is a probability measure λ on S such that

∀x ∈ C(S) ‖ux‖ ≤√

kG‖u‖(∫

|x|2dλ

)1/2

.

(ii) For any finite set (x1, . . . , xn) in C(S) we have

(∑‖uxj‖2

)1/2≤

√kG‖u‖

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

.

Moreover√

kG is the best possible constant in either (i) or (ii).

The equivalence of (i) and (ii) is a particular case of Proposition 4.5. By Cauchy-Schwarz,Theorem 5.2 obviously implies Theorem 5.1. Conversely, applying Theorem 5.1 to the form (x, y) 7→〈u(x), u(y)〉 and taking xj = yj in (5.1) we obtain Theorem 5.2.

The next Lemma will serve to minorize kG (and hence KG).

Lemma 5.3. Let (Ω, µ) be an arbitrary measure space. Consider gj ∈ L1(µ) and xj ∈ L∞(µ)(1 ≤ j ≤ N) and positive numbers a, b. Assume that (gj) and (xj) are biorthogonal (i.e. 〈gi, xj〉 = 0if i 6= j and = 1 otherwise) and such that

∀(αj) ∈ KN a(

∑|αj |2)1/2 ≥ ‖

∑αjgj‖1 and

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

≤ b√

N.

Then b−2a−2 ≤ kG (and a fortiori b−2a−2 ≤ KG).

Proof. Let H = ℓN2 . Let u : L∞ → H be defined by u(x) =

∑〈x, gi〉ei. Our assumptions imply∑ ‖u(xj)‖2 = N , and ‖u‖ ≤ a. By (ii) in Theorem 5.2 we obtain b−1a−1 ≤ √kG.

It turns out that Theorem 5.2 can be proved directly as a consequence of Khintchine’s inequalityin L1 or its Gaussian analogue. The Gaussian case yields the best possible k.

Let (εj) be an i.i.d. sequence of ±1-valued random variables with P(εj = ±1) = 1/2. Thenfor any scalar sequence (αj) we have

(5.4)1√2

(∑|αj |2

)1/2≤

∥∥∥∑

αjεj

∥∥∥1≤

(∑|αj |2

)1/2.

Of course the second inequality is trivial since ‖∑εjαj‖2 = (

∑ |αj |2)1/2.The Gaussian analogue of this equivalence is the following equality:

(5.5)∥∥∥∑

αjgj

∥∥∥1

= ‖g‖1

(∑|αj |2

)1/2.

(Recall that (gj) is an i.i.d. sequence of copies of g.) Note that each of these assertions correspondsto an embedding of ℓ2 into L1(Ω, P), an isomorphic one in case of (5.4), isometric in case of (5.5).

A typical application of (5.5) is as follows: let v : H → L1(Ω′, µ′) be a bounded operator. Then

for any finite sequence (yj) in H

(5.6) ‖g‖1

∥∥∥∥(∑

|v(yj)|2)1/2

∥∥∥∥1

≤ ‖v‖(∑

‖yj‖2)1/2

.

21

Indeed, we have by (5.5)′

‖g‖1

∥∥∥∥(∑

|v(yj)|2)1/2

∥∥∥∥1

=∥∥∥∑

gjv(yj)∥∥∥

L1(P×µ′)=

∫ ∥∥∥v(∑

gj(ω)yj

)∥∥∥1dP(ω)

≤ ‖v‖(∫ ∥∥∥

∑gj(ω)yj

∥∥∥2dP

)= ‖v‖

(∑‖yj‖2

)1/2.

Consider now the adjoint v∗ : L∞(µ′) → H∗. Let u = v∗. Note ‖u‖ = ‖v‖. By duality, it is easyto deduce from (5.6) that for any finite subset (x1, . . . , xn) in L∞(µ′) we have

(5.7)(∑

‖uxj‖2)1/2

≤ ‖u‖‖g‖−11

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

.

This leads us to:

Proof of Theorem 5.2 (and Theorem 5.1). By the preceding observations, it remains to prove (ii)and justify the value of kG. Here again by suitable approximation (L∞-spaces) we may reduce tothe case when S is a finite set. Then C(S) = ℓ∞(S) and so that if we apply (5.7) to u we obtain(ii) with

√k ≤ ‖g‖−1

1 . This shows that kG ≤ ‖g‖−21 . To show that kG ≥ ‖g‖−2

1 , we will use Lemma5.3. Let xj = cNN1/2gj(

∑ |gj |2)−1/2 with cN adjusted so that (xj) is biorthogonal to (gj), i.e.we set N−1/2c−1

N =∫|g1|2(

∑ |gj |2)−1/2. Note∫|gj |2(

∑ |gj |2)−1/2 =∫|g1|2(

∑ |gj |2)−1/2 for anyj and hence c−1

N = N−1/2∑

j

∫|gj |2(

∑ |gj |2)−1/2 = N−1/2∫

(∑ |gj |2)1/2. Thus by the strong law

of large numbers (actually here the weak law suffices), since E|gj |2 = 1 c−1N → 1 when N → ∞.

Then by Lemma 5.3 (recall (5.5)) we find c−1N ‖g‖−1

1 ≤√

kG. Thus letting N → ∞ we obtain‖g‖−1

1 ≤ √kG.

6. Banach spaces satisfying GT and operator ideals

It is natural to try to extend GT to Banach spaces other than C(S) or L∞ (or their duals). Sinceany Banach space is isometric to a subspace of C(S) for some compact S, one can phrase thequestion like this: What are the pairs of subspaces X ⊂ C(S), Y ⊂ C(T ) such that any boundedbilinear form ϕ : X × Y → K satisfies the conclusion of Theorem 2.1? Equivalently, this propertymeans that ‖ · ‖H and ‖ ‖∧ are equivalent on X ⊗ Y . This reformulation shows that the propertydoes not depend on the choice of the embeddings X ⊂ C(S), Y ⊂ C(T ). Since the reference [115]contains a lot of information on this question, we will merely briefline outline what is known andpoint to a few more recent sources.

Definition 6.1. (i) A pair of (real or complex) Banach space (X, Y ) will be called a GT-pair if anybounded bilinear form ϕ : X × Y → K satisfies the conclusion of Theorem 2.1, with say S, T equalrespectively to the weak∗ unit balls of X∗, Y ∗. Equivalently, this means that there is a constant Csuch that

∀t ∈ X ⊗ Y ‖t‖∧ ≤ C‖t‖H .

(ii) A Banach space X is called a GT-space (in the sense of [115]) if (X∗, C(T )) is a GT-pair forany compact set T .

GT tells us that the pairs (C(S), C(T )) or (L∞(µ), L∞(µ′)) are GT-pairs, and that all abstractL1-spaces (this includes C(S)∗) are GT-spaces.

22

Remark 6.2. If (X, Y ) is a GT-pair and we have isomorphic embeddings X ⊂ X1 and Y ⊂ Y1, witharbitrary X1, Y1, then any bounded bilinear form on X × Y extends to one on X1 × Y1.

Let us say that a pair (X, Y ) is a “Hilbertian pair” if every bounded linear operator from X toY ∗ factors through a Hilbert space. Equivalently this means that X⊗Y and X⊗H′Y coincide withequivalent norms. It is easy to see that (X, Y ) is a GT-pair iff it is a Hilbertian pair and moreovereach operator from X to ℓ2 and from Y to ℓ2 is 2-summing.

It is known (see [115]) that (X, C(T )) is a GT-pair for all T iff X∗ is a GT-space. See [78]for a discussion of GT-pairs. The main examples are pairs (X, Y ) such that X⊥ ⊂ C(S)∗ andY ⊥ ⊂ C(T )∗ are both reflexive (proved independently by Kisliakov and the author in 1976), alsothe pairs (X, Y ) with X = Y = A(D) (disc algebra) and X = Y = H∞ (proved by Bourgain in1981 and 1984). In particular, Bourgain’s result shows that if (using boundary values) we view H∞

over the disc as isometrically embedded in the space L∞ over the unit circle, then any boundedbilinear form on H∞ ×H∞ extends to a bounded bilinear form on L∞ ×L∞. See [116] for a proofthat the pair (J, J) is a Hilbertian pair, when J is the classical James space of codimension 1 in itsbidual.

On a more “abstract” level, if X∗ and Y ∗ are both GT-spaces of cotype 2 (resp. both of cotype2) and if one of them has the approximation property then (X, Y ) is a GT pair (resp. a Hilbertianpair). See Definition 9.6 below for the precise definition of ”cotype 2”. This “abstract” form of GTwas crucially used in the author’s construction of infinite dimensional Banach spaces X such that

X∧⊗ X = X

∨⊗ X, i.e. ‖ ‖∧ and ‖ ‖∨ are equivalent norms on X ⊗ X. See [115] for more precise

references on all this.As for more recent results on the same direction see [59] for examples of pairs of infinite dimen-

sional Banach spaces X, Y such that any compact operator from X to Y is nuclear. Note that thereis still no nice characterization of Hilbertian pairs. See [88] for a counterexample to a conjecturein [115] on this.

We refer the reader to [79] and [33] for more recent updates on GT in connection with Banachspaces of analytic functions and uniform algebras.

Grothendieck’s work had a major impact on Kwapien’s and Pietsch’s theory of operator ideals[86, 111]. The latter itself had a huge impact on the development of the Geometry of Banach spacesin the last 4 decades. We refer the reader to [143, 24] for more on this development.

7. Non-commutative GT

Grothendieck himself conjectured ([36, p. 73]) a non-commutative version of Theorem 2.1. Thiswas proved in [112] with an additional approximation assumption, and in [45] in general. Actually,a posteriori, by [45], the assumption needed in [112] always holds. Incidentally it is amusing tonote that Grothendieck overlooked the difficulty related to the approximation property, since heasserts without proof that the problem can be reduced to finite dimensional C∗-algebras.

In the optimal form proved in [45], the result is as follows.

Theorem 7.1. Let A, B be C∗-algebras. Then for any bounded bilinear form ϕ : A×B → C thereare states f1, f2 on A, g1, g2 on B such that

(7.1) ∀(x, y) ∈ A × B |ϕ(x, y)| ≤ ‖ϕ‖(f1(x∗x) + f2(xx∗))1/2(g1(yy∗) + g2(y

∗y))1/2.

Actually, just as (2.1) and (2.3), (7.1) is equivalent to the following inequality valid for all finitesets (xj , yj) 1 ≤ j ≤ n in A × B

(7.2)∣∣∣∑

ϕ(xj , yj)∣∣∣ ≤ ‖ϕ‖

(∥∥∥∑

x∗jxj

∥∥∥ +∥∥∥∑

xjx∗j

∥∥∥)1/2 (∥∥∥

∑y∗j yj

∥∥∥ +∥∥∥∑

yjy∗j

∥∥∥)1/2

.

23

A fortiori we have(7.3)∣∣∣

∑ϕ(xj , yj)

∣∣∣ ≤ 2‖ϕ‖max

∥∥∥∑

x∗jxj

∥∥∥1/2

,∥∥∥∑

xjx∗j

∥∥∥1/2

max

∥∥∥∑

y∗j yj

∥∥∥1/2

,∥∥∥∑

yjy∗j

∥∥∥1/2

.

For further reference, we state the following obvious consequence:

Corollary 7.2. Let A, B be C∗-algebras and let u : A → H and v : B → H be bounded linearoperators into a Hilbert space H. Then for all finite sets (xj , yj) 1 ≤ j ≤ n in A × B we have(7.4)

|∑

〈u(xj), v(yj)〉| ≤ 2‖u‖‖v‖max‖∑

x∗jxj‖1/2, ‖

∑xjx

∗j‖1/2max‖

∑y∗j yj‖1/2, ‖

∑yjy

∗j ‖1/2.

It was proved in [47] that the constant 1 is best possible in either (7.1) or (7.2). Note howeverthat “constant = 1” is a bit misleading since if A, B are both commutative, (7.1) or (7.2) yield(2.1) or (2.3) with K = 2.

Lemma 7.3. Consider C∗-algebras A, B and subspaces E ⊂ A and F ⊂ B. Let u : E → F ∗ alinear map with associated bilinear form ϕ(x, y) = 〈ux, y〉 on E × F . Assume that there are statesf1, f2 on A, g1, g2 on B such that

∀(x, y) ∈ E × F |〈ux, y〉| ≤ (f1(xx∗) + f2(x∗x))1/2(g1(y

∗y) + g2(yy∗))1/2.

Then ϕ : E × F → C admits a bounded bilinear extension ϕ : A × B → C that can be decomposedas a sum

ϕ = ϕ1 + ϕ2 + ϕ3 + ϕ4

where ϕj : A× B → C are bounded bilinear forms satisfying the following: For all (x, y) in A × B

|ϕ1(x, y)| ≤ (f1(xx∗)g1(y∗y))1/2(6.3)1

|ϕ2(x, y)| ≤ (f2(x∗x)g2(yy∗))1/2(6.3)2

|ϕ3(x, y)| ≤ (f1(xx∗)g2(yy∗))1/2(6.3)3

|ϕ4(x, y)| ≤ (f2(x∗x)g1(y

∗y))1/2.(6.3)4

A fortiori, we can write u = u1 + u2 + u3 + u4 where uj : E → F ∗ satisfies the same bound as ϕj

for (x, y) in E × F .

Proof. Let H1 and H2 (resp. K1 and K2) be the Hilbert spaces obtained from A (resp. B) withrespect to the (non Hausdorff) inner products 〈a, b〉H1 = f1(ab∗) and 〈a, b〉H2 = f2(b

∗a) (resp.〈a, b〉K1 = g1(b

∗a) and 〈a, b〉K2 = g2(ab∗)). Then our assumption can be rewritten as |〈ux, y〉| ≤‖(x, x)‖H1⊕H2‖(y, y)‖K1⊕K2 . Therefore using the orthogonal projection from H1⊕H2 onto span[(x⊕x) | x ∈ E] and similarly for K1⊕K2, we find an operator U : H1⊕H2 → (K1⊕K2)

∗ with ‖U‖ ≤ 1such that

(7.6) ∀(x, y) ∈ E × F 〈ux, y〉 = 〈U(x ⊕ x), (y ⊕ y)〉.Clearly we have contractive linear “inclusions”

A ⊂ H1, A ⊂ H2, B ⊂ K1, B ⊂ K2

so that the bilinear forms ϕ1, ϕ2, ϕ3, ϕ4 defined on A × B by the identity

〈U(x1 ⊕ x2), (y1 ⊕ y2)〉 = ϕ1(x1, y1) + ϕ2(x2, y2) + ϕ3(x1, y2) + ϕ4(x2, y1)

must satisfy the inequalities (6.3)1 and after. By (7.6), we have (ϕ1 + ϕ2 + ϕ3 + ϕ4)|E×F = ϕ.Equivalently, if uj : E → F ∗ are the linear maps associated to ϕj , we have u = u1+u2+u3+u4.

24

8. Non-commutative “little GT”

The next result was first proved in [112] with a worse constant. Haagerup [45] obtained the constant1, that was shown to be optimal in [47].

Theorem 8.1. Let A be a C∗-algebra, H a Hilbert space. Then for any bounded linear mapu : A → H there are states f1, f2 on A such that

∀x ∈ A ‖ux‖ ≤ ‖u‖(f1(x∗x) + f2(xx∗))1/2.

Equivalently, for any finite set (x1, . . . , xn) in A we have

(8.1)(∑

‖uxj‖2)1/2

≤ ‖u‖(∥∥∥

∑x∗

jxj

∥∥∥ +∥∥∥∑

xjx∗j

∥∥∥)1/2

,

and a fortiori

(8.2)(∑

‖uxj‖2)1/2

≤√

2‖u‖max

∥∥∥∑

x∗jxj

∥∥∥1/2

,∥∥∥∑

xjx∗j

∥∥∥1/2

.

Proof. This can be deduced from Theorem 7.1 (or Corollary 7.2 exactly as in §5, by consideringthe bounded bilinear (actually sesquilinear) form ϕ(x, y) = 〈ux, uy〉.

Note that ‖(∑x∗jxj)

1/2‖ = ‖∑x∗

jxj‖1/2. At first sight, the reader may view ‖(∑x∗jxj)

1/2‖A

as the natural generalization of the norm ‖(∑ |xj |2)1/2‖∞ appearing in Theorem 2.1 in case A iscommutative. There is however a major difference: if A1, A2 are commutative C∗-algebras, thenfor any bounded u : A1 → A2 we have for any x1, . . . , xn in A1

(8.3)

∥∥∥∥(∑

u(xj)∗u(xj)

)1/2∥∥∥∥ ≤ ‖u‖

∥∥∥∥(∑

x∗jxj

)1/2∥∥∥∥ .

The simplest way to check this is to observe that

(8.4)

∥∥∥∥(∑

|xj |2)1/2

∥∥∥∥∞

= sup∥∥∥

∑αjxj

∥∥∥∞

∣∣∣ αj ∈ K

∑|αj |2 ≤ 1

.

Indeed, (8.4) shows that the seemingly non-linear expression ‖(∑ |xj |2)1/2‖∞ can be suitably “lin-earized” so that (8.3) holds. But this is no longer true when A1 is not-commutative. In fact, letA1 = A2 = Mn and u : Mn → Mn be the transposition of n × n matrices, then if we set xj = e1j ,we have uxj = ej1 and we find ‖(∑ x∗

jxj)1/2‖ = 1 but ‖(∑(uxj)

∗(uxj))1/2‖ =

√n. This shows that

there is no non-commutative analogue of the “linearization” (8.4) (even as a two-sided equivalence).The “right” substitute seems to be the following corollary.

Notation. Let x1, . . . , xn be a finite set in a C∗-algebra A. We denote

(8.5) ‖(xj)‖C =

∥∥∥∥(∑

x∗jxj

)1/2∥∥∥∥ , ‖(xj)‖R =

∥∥∥∥(∑

xjx∗)1/2

∥∥∥∥ , and :

(8.6) ‖(xj)‖RC = max‖(xj)‖R, ‖(xj)‖C.

It is easy to check that ‖.‖C , ‖.‖R and hence also ‖.‖RC are norms on the space of finitely supportedsequences in A.

25

Corollary 8.2. Let A1, A2 be C∗-algebras. Let u : A1 → A2 be a bounded linear map. Then forany finite set x1, . . . , xn in A1 we have

(8.7) ‖(uxj)‖RC ≤√

2‖u‖‖(xj)‖RC .

Proof. Let ξ be any state on A2. Let L2(ξ) be the Hilbert space obtained from A2 equipped withthe inner product 〈a, b〉 = ξ(b∗a). (This is the so-called “GNS-construction”). We then have acanonical inclusion jξ : A2 → L2(ξ). Then (8.1) applied to the composition jξu yields

(∑ξ((uxj)

∗(uxj)))1/2

≤ ‖u‖√

2‖(xj)‖RC .

Taking the supremum over all ξ we find

‖(uxj)‖C ≤ ‖u‖√

2‖(xj)‖RC .

Similarly (taking 〈a, b〉 = ξ(ab∗) instead) we find

‖(uxj)‖R ≤ ‖u‖√

2‖(xj)‖RC .

Remark 8.3. The problem of extending the non-commutative GT from C∗-algebras to JB∗-tripleswas considered notably by Barton and Friedman around 1987, but seems to be still incomplete, see[108, 109] for a discussion and more precise references.

9. Non-commutative Khintchine inequality

In analogy with (8.4), it is natural to expect that there is a “linearization” of [(xj)]RC that is behind(8.7). This is one of the applications of the Khintchine inequality in non-commutative L1, i.e. thenon-commutative version of (5.4) and (5.5).

Recall first that by “non-commutative L1-space,” one usually means a Banach space X such thatits dual X∗ is a von Neumann algebra. (We could equivalently say “such that X∗ is (isometric to) aC∗-algebra” because a C∗-algebra that is also a dual must be a von Neumann algebra isometrically.)Since the commutative case is almost always the basic example for the theory it seems silly toexclude it, so we will say instead that X is a generalized (possibly non-commutative) L1-space.When M is a commutative von Neumann algebra, we have M ≃ L∞(Ω, µ) isometrically for someabstract measure space (Ω, µ) and hence if M = X∗, X ≃ L1(Ω, µ).

For example if M = B(H) then the usual trace x 7→ tr(x) is semi-finite and L1(M, tr) can beidentified with the trace class (usually denoted by S1(H) or simply S1) that is formed of all thecompact operators x : H → H such that tr(|x|) < ∞ with ‖x‖S1 = tr(|x|). (Here of course byconvention |x| = (x∗x)1/2.) It is classical that S1(H)∗ ≃ B(H) isometrically.

When M is a non-commutative von Neumann algebra the measure µ is replaced by a trace,i.e. an additive, positively homogeneous functional τ : M+ → [0,∞], such that τ(xy) = τ(yx)for all x, y in M+. The trace τ is usually assumed “normal” (this is equivalent to σ-additivity, i.e.∑

τ(Pi) = τ(∑

Pi) for any family (Pi) of mutually orthogonal self-adjoint projections Pi in M) and“semi-finite” (i.e. the set of x ∈ M+ such that τ(x) < ∞ generates M as a von Neumann algebra).One also assumes τ “faithful” (i.e. τ(x) = 0 ⇒ x = 0). One can then mimic the construction ofL1(Ω, µ) and construct the space X = L1(M, τ) in such a way that L1(M, τ)∗ = M isometrically.This analogy explains why one often sets L∞(M, τ) = M and one denotes by ‖ ‖∞ and ‖ ‖1

26

respectively the norms in M and L1(M, τ). Consider now y1, . . . , yn ∈ L1(M, τ). The followingequalities are easy to check:

(9.1) τ

((∑y∗j yj

)1/2)

= sup∣∣∣

∑τ(xjyj)

∣∣∣∣∣∣ ‖(xj)‖R ≤ 1

(9.2) τ

((∑yjy

∗j

)1/2)

= sup∣∣∣

∑τ(xjyj)

∣∣∣∣∣∣ ‖(xj)‖C ≤ 1

.

In other words the norm (xj) 7→ ‖(∑ x∗jxj)

1/2‖∞ = ‖(xj)‖C is naturally in duality with the norm

(yj) 7→ ‖(∑ yjy∗j )

1/2‖1, and similarly for (xj) 7→ ‖(xj)‖R. Incidentally the last two equalities show

that (yj) 7→ ‖(∑ yjy∗j )

1/2‖1 and (yj) 7→ ‖(∑ y∗j yj)1/2‖1 satisfy the triangle inequality and hence

are indeed norms.The von Neumann algebras M that admit a trace τ as above are called “semi-finite,” but there

are many examples that are not semi-finite. To cover that case too in the sequel we make thefollowing convention. If X is any generalized (possibly non-commutative) L1-space, with M = X∗

possibly non-semifinite, then for any (y1, . . . , yn) in X we set by definition

‖(yj)‖1,R = sup∣∣∣

∑〈xj , yj〉

∣∣∣∣∣∣ xj ∈ M, ‖(xj)‖C ≤ 1

‖(yj)‖1,C = sup∣∣∣

∑〈xj , yj〉

∣∣∣∣∣∣ xj ∈ M, ‖(xj)‖R ≤ 1

.

Here of course 〈·, ·〉 denotes the duality between X and M = X∗. Since M admits a unique predual(up to isometry) it is customary to set M∗ = X.

Notation. For (y1, . . . , yn) is M∗ we set

|||(yj)|||1 = inf‖(y′j)‖1,R + ‖(y′′j )‖1,C

where the infimum runs over all possible decompositions of the form yj = y′j + y′′j , j = 1, . . . , n. Byan elementary duality argument, one deduces form (9.1) and (9.2) that for all (y1, . . . , yn) is M∗

(9.3) |||(yj)|||1 = sup∣∣∣

∑〈xj , yj〉

∣∣∣∣∣∣ xj ∈ M, max‖(xj)‖C , ‖(xj)‖R ≤ 1

.

We will denote by (gRj ) (resp. (gC

j )) an independent sequence of real (resp. complex) valued Gaussianrandom variables with mean zero and L2-norm 1. We also denote by (sj) an independent sequenceof complex valued variables, each one uniformly distributed over the unit circle T. This is thecomplex analogue (“Steinhaus variables”) of the sequence (εj). We can now state the Khintchineinequality for (possibly) non-commutative L1-spaces, and its Gaussian counterpart:

Theorem 9.1. There are constants c1, cC1 , c1 and cC

1 such that for any M and any finite set(y1, . . . , yn) in M∗ we have (recall ‖ · ‖1 = ‖ ‖M∗)

1

c1|||(yj)|||1 ≤

∫ ∥∥∥∑

εj(ω)yj

∥∥∥1dP(ω) ≤ |||(yj)|||1(9.4)

1

cC1

|||(yj)|||1 ≤∫ ∥∥∥

∑sj(ω)yj

∥∥∥1dP(ω) ≤ |||(yj)|||1(9.5)

1

c1|||(yj)|||1 ≤

∫ ∥∥∥∑

gRj (ω)yj

∥∥∥1dP(ω) ≤ |||(yj)|||(9.6)

1

cC1

|||(yj)|||1 ≤∫ ∥∥∥

∑gCj (ω)yj

∥∥∥1dP(ω) ≤ |||(yj)|||1.(9.7)

27

This result was first proved in [94]. Actually [94] contains two proofs of it, one that derivesit from the “non-commutative little GT” and the so-called cotype 2 property of M∗, another onebased on the factorization of functions in the Hardy space of M∗-valued functions H1(M∗). With alittle polish (see [118, p. 347 ]), the second proof yields (9.5) with cC

1 = 2, and hence cC1 = 2 by an

easy central limit argument. More recently, Haagerup and Musat ([48]) found a proof of (9.6) and(9.7) with cC

1 = cC1 =

√2, and by [47] these are the best constants here (see Theorem 11.4 below).

They also proved that c1 ≤√

3 (and hence c1 ≤√

3 by the central limit theorem) but the bestvalues of c1 and c1 remain apparently unknown.

To prove (9.4), we will use the following.

Lemma 9.2 ([112, 45]). Let M be a C∗-algebra. Consider x1, . . . , xn ∈ M . Let S =∑

εkxk andlet S =

∑skxk. Assuming x∗

k = xk, we have

(9.8)

∥∥∥∥∥

(∫S4 dP

)1/4∥∥∥∥∥ ≤ 31/4

∥∥∥∥(∑

x2k

)1/2∥∥∥∥ .

No longer assuming (xk) self-adjoint, we set T =

(0 S

S∗ 0

)so that T = T ∗. We have then:

(9.9)

∥∥∥∥∥

(∫T 4 dP

)1/4∥∥∥∥∥ ≤ 21/4‖(xk)‖RC .

Remark. By the central limit theorem, 31/4 and 21/4 are clearly the best constants here, becauseE|gR|4 = 3 and E|gC|4 = 2.

Proof of Theorem 9.1. By duality (9.4) is clearly equivalent to: ∀n∀x1, . . . , xn ∈ M

(9.10) [(xk)] ≤ c1‖(xk)‖RC

where

[(xk)]def= inf

‖Φ‖L∞(P;M) |

∫εkΦ dP = xk, k = 1, . . . , n

.

The property∫

εkΦ dP = xk (k = 1, . . . , n) can be reformulated as Φ =∑

εkxk + Φ′ with Φ′ ∈[εk]

⊥ ⊗ M . Note for further use that we have

(9.11)

∥∥∥∥(∑

x∗kxk

)1/2∥∥∥∥ ≤

∥∥∥∥∥

(∫Φ∗Φ dP

)1/2∥∥∥∥∥

M

.

Indeed (9.11) is immediate by orthogonality because

∑x∗

kxk ≤∑

x∗kxk +

∫Φ′∗Φ′ dP =

∫Φ∗Φ dP.

By an obvious iteration, it suffices to show the following.Claim: If ‖(xk)‖RC < 1, then there is Φ in L∞(P; M) with ‖Φ‖L∞(P;M) ≤ c1/2 such that if

xkdef=

∫εkΦ dP we have ‖(xk − xk)‖RC < 1

2 .

Since the idea is crystal clear in this case, we will first prove this claim with c1 = 4√

3 and onlyafter that indicate the refinement that yields c1 =

√3. It is easy to reduce the claim to the case

when xk = x∗k (and we will find xk also self-adjoint). Let S =

∑εkxk. We set

(9.12) Φ = S1|S|≤2√

3.

28

Here we use a rather bold notation: we denote by 1|S|≤λ (resp. 1|S|>λ) the spectral projectioncorresponding to the set [0, λ] (resp. (λ,∞)) in the spectral decomposition of the (self-adjoint)operator |S|. Note Φ = Φ∗ (since we assume S = S∗). Let F = S − Φ = S1|S|>2

√3. By (9.8)

∥∥∥∥∥

(∫S4 dP

)1/2∥∥∥∥∥ ≤ 31/2

∥∥∥∑

x2k

∥∥∥ < 31/2.

A fortiori ∥∥∥∥∥

(∫F 4 dP

)1/2∥∥∥∥∥ < 31/2

but since F 4 ≥ F 2(2√

3)2 this implies∥∥∥∥∥

(∫F 2 dP

)1/2∥∥∥∥∥ < 1/2.

Now if xk =∫

Φεk dP, i.e. if Φ =∑

εkxk +Φ′ with Φ′ ∈ [εk]⊥⊗M we find F =

∑εk(xk − xk)−Φ′

and hence by (9.11) applied with F in place of Φ (note Φ, F, xk are all self-adjoint)

‖(xk − xk)‖RC = ‖(xk − xk)‖C ≤∥∥∥∥∥

(∫F 2 dP

)1/2∥∥∥∥∥ < 1/2,

which proves our claim.To obtain this with c1 =

√3 instead of 4

√3 one needs to do the truncation (9.12) in a slightly

less brutal way: just setΦ = S1|S|≤c + c1S>c − c1S<−c

where c =√

3/2. A simple calculation then shows that F = S − Φ = fc(S) where |fc(t)| =1(|t|>c)(|t| − c). Since |fc(t)|2 ≤ (4c)−2t4 for all t > 0 we have

(∫F 2 dP

)1/2

≤ (2√

3)−1

(∫S4 dP

)1/2

and hence by (9.8) ∥∥∥∥∥

(∫F 2 dP

)1/2∥∥∥∥∥ < 1/2

so that we obtain the claim as before using (9.11), whence (9.4) with c1 =√

3. The proof of(9.5) and (9.7) (i.e. the complex cases) with the constant cC

1 = cC1 =

√2 follows entirely parallel

steps, but the reduction to the self-adjoint case is not so easy, so the calculations are slightly moreinvolved. The relevant analogue of (9.8) in that case is (9.9).

Remark 9.3. In [54] (see also [122, 125]) versions of the non-commutative Khintchine’s inequalityin L4 are proved for the sequences of functions of the form eint | n ∈ Λ in L2(T) that satisfy theZ(2)-property in the sense that there is a constant C such that

supn∈Z

card(k, m) ∈ Z2 | k 6= m, k − m ∈ Λ ≤ C.

It is easy to see that the method from [49] (to deduce the L1-case from the L4-one) described inthe proof of Theorem 9.1 applies equally well to the Z(2)-subsets of Z or of arbitrary (possiblynon-commutative) discrete groups.

29

Although we do not want to go too far in that direction, we cannot end this section withoutdescribing the non-commutative Khintchine inequality for values of p other than p = 1.

Consider a generalized (possibly non-commutative) measure space (M, τ) (recall we assume τsemi-finite). The space Lp(M, τ) or simply Lp(τ) can then be described as a complex interpolationspace, i.e. we can use as a definition the isometric identity (here 1 < p < ∞)

Lp(τ) = (L∞(τ), L1(τ)) 1p.

The case p = 2 is of course obvious: for any x1, . . . , xn in L2(τ) we have

(∫ ∥∥∥∑

εkxk

∥∥∥2

L2(τ)dP

)1/2

=(∑

‖xk‖2L2(τ)

)1/2,

but the analogous 2-sided inequality for

(∫ ∥∥∥∑

εkxk

∥∥∥p

Lp(τ)dP

)1/p

with p 6= 2 is not so easy to describe, in part because it depends very much whether p < 2 or p > 2(and moreover the case 0 < p < 1 is still open!).

Assume 1 ≤ p < ∞ and xj ∈ Lp(τ) (j = 1, . . . , n). We will use the notation

‖(xj)‖p,R =

∥∥∥∥(∑

xjx∗j

)1/2∥∥∥∥

p

and ‖(xj)‖p,C =

∥∥∥∥(∑

x∗jxj

)1/2∥∥∥∥

p

.

Remark 9.4. Let x = (xj). Let R(x) =∑

e1j ⊗xj (resp. C(x) =∑

ej1 ⊗xj) denote the row (resp.column) matrix with entries enumerated as x1, · · · , xn. We may clearly view R(x) and C(x) aselements of Lp(Mn⊗M, tr⊗τ). Computing |R(x)∗| = (R(x)R(x)∗)1/2 and |C(x)| = (C(x)∗C(x))1/2,we find

‖(xj)‖p,R = ‖R(x)‖Lp(Mn⊗M,tr⊗τ) and ‖(xj)‖p,C = ‖C(x)‖Lp(Mn⊗M,tr⊗τ).

This shows in particular that (xj) 7→ ‖(xj)‖p,R and (xj) 7→ ‖(xj)‖p,C are norms on the space offinitely supported sequences in Lp(τ). Whenever p 6= 2, the latter norms are different. Actually,they are not even equivalent with constants independent of n. Indeed a simple example is providedby the choice of M = Mn (or M = B(ℓ2)) equipped with the usual trace and xj = ej1 (j = 1, . . . , n).We have then ‖(xj)‖p,R = n1/p and ‖(xj)‖p,C = ‖(txj)‖p,R = n1/2.

Now assume 1 ≤ p ≤ 2. In that case we define

|||(xj)|||p = inf‖(y′j)‖p,R + ‖(y′′j )‖p,C

where the infimum runs over all possible decompositions of (xj) of the form xj = y′j + y′′j , y′j , y′′j ∈

Lp(τ). In sharp contrast, when 2 ≤ p < ∞ we define

|||(xj)|||p = max‖(xj)‖p,R, ‖(xj)‖p,C.

The non-commutative Khintchine inequalities that follow are due to Lust–Piquard ([92]) in case1 < p 6= 2 < ∞.

30

Theorem 9.5 (Non-commutative Khintchine in Lp(τ)). (i) If 1 ≤ p ≤ 2 there is a constantcp > 0 such that for any finite set (x1, . . . , xn) in Lp(τ) we have

(9.13)1

cp|||(xj)|||p ≤

(∫ ∥∥∥∑

εjxj

∥∥∥p

pdP

)1/p

≤ |||(xj)|||p.

(ii) If 2 ≤ p < ∞, there is a constant bp such that for any (x1, . . . , xn) in Lp(τ)

(9.14) |||(xj)|||p ≤(∫ ∥∥∥

∑εjxj

∥∥∥p

pdP

)1/p

≤ bp|||(xj)|||p.

Moreover, (9.13) and (9.14) also hold for the sequences (sj), (gRj ) or (gC

j ). We will denote the

corresponding constants respectively on one hand by cCp , cp and cC

p , and on the other hand by bCp ,

bp and bCp .

Definition 9.6. A Banach space B is called “of cotype 2” if there is a constant C > 0 such thatfor all finite sets (xj) in B we have

(∑‖xj‖2

)1/2≤ C

∥∥∥∑

εjxj

∥∥∥L2(B)

.

It is called “of type 2” if the inequality is in the reverse direction.

From the preceding Theorem, one recovers easily the known result from [142] that Lp(τ) is ofcotype 2 (resp. type 2) whenever 1 ≤ p ≤ 2 (resp. 2 ≤ p < ∞).Remarks:

(i) In the commutative case, the preceding result is easy to obtain by integration from the classicalKhintchine inequalities, for which the best constants are known (cf. [140], [42], [136]), they are

respectively cp = 21p− 1

2 for 1 ≤ p ≤ p0 and cp = ‖gR1 ‖p = for p0 ≤ p < ∞, where 1 < p0 < 2 is

determined by the equality

21

p0− 1

2 = ‖gR1 ‖p0 .

Numerical calculation yields p0 = 1.8474...!

(ii) In the non-commutative case, say when Lp(τ) = Sp (Schatten p-class) not much is known onthe best constants, except for the case p = 1 (see the next section for that). Note howeverthat if p is an even integer the (best possible) type 2 constant of Lp(τ) was computed alreadyby Tomczak–Jaegermann in [142]. One can deduce from this (see [123, p. 193]) that theconstant bp in (9.14) is O(

√p) when p → ∞, which is at least the optimal order of growth, see

also [19] for exact values of bp for p an even integer. Note however that, by [94], cp remainsbounded when 1 ≤ p ≤ 2.

(iii) Whether (9.13) holds for 0 < p < 1 is an open problem, although many natural consequencesof this have been verified (see [154, 125]). See [71] for somewhat related topics.

The next result is meant to illustrate the power of (9.13) and (9.14). We first introduce aproblem. Consider a complex matrix [aij ] and let (εij) denote independent random signs so thatεij | i, j ≥ 0 is an i.i.d. family of ±1-valued variables with P(εij = ±1) = 1/2.

31

Problem. Fix 0 < p < ∞. Characterize the matrices [aij ] such that

[εijaij ] ∈ Sp

for almost all choices of signs εij .Here is the solution:

Corollary 9.7. When 2 ≤ p < ∞, [εij , aij ] ∈ Sp a.s. iff we have both

∑i

(∑j|aij |2

)p/2< ∞ and

∑j

(∑i|aij |2

)p/2< ∞.

When 0 < p ≤ 2, [εijaij ] ∈ Sp a.s. iff there is a decomposition aij = a′ij+a′′ij with∑

i(∑

j |a′ij |2)p/2 <

∞ and∑

j(∑

i |a′′ij |2)p/2 < ∞.

This was proved in [92] for 1 < p < ∞ ([94] for p = 1) as an immediate consequence of Theorem9.5 (or Theorem 9.1) applied to the series

∑εijaijeij , together with known general facts about

random series of vectors on a Banach space (cf. e.g. [69, 57]). The case 0 < p < 1 was recentlyobtained in [125].

Corollary 9.8 (Non-commutative Marcinkiewicz–Zygmund). Let Lp(M1, τ1), Lp(M2, τ2) be gener-alized (possibly non-commutative) Lp-spaces, 1 ≤ p < ∞. Then there is a constant K(p) such thatany bounded linear map u : Lp(τ1) → Lp(τ2) satisfies the following inequality: for any (x1, . . . , xn)in Lp(τ1)

|||(uxj)|||p ≤ K(p)‖u‖ |||(xj)|||p.We have K(p) ≤ cp if 1 ≤ p ≤ 2 and K(p) ≤ bp if p ≥ 2.

Proof. Since we have trivially for any fixed εj = ±1

∥∥∥∑

εjuxj

∥∥∥p≤ ‖u‖

∥∥∥∑

εjxj

∥∥∥p,

the result is immediate from Theorem 9.5.

Remark. The case 0 < p < 1 is still open.

Remark. The preceding corollary is false (assuming p 6= 2) if one replaces ||| |||p by either ‖(·)‖p,C

or ‖(·)‖p,R. Indeed the transposition u : x → tx provides a counterexample. See Remark 9.4.

10. Maurey factorization

Let (Ω, µ) be a measure space. In his landmark paper [134] Rosenthal made the following observa-tion: let u : L∞(µ) → X be a operator into a Banach space X that is 2-summing with π2(u) ≤ c, orequivalently satisfies (4.1). If u∗(X∗) ⊂ L1(µ) ⊂ L∞(µ)∗, then the probability λ associated to (4.1)that is a priori in L∞(µ)∗+ can be chosen absolutely continuous with respect to µ, so we can takeλ = f ·µ with f a probability density on Ω. Then we find ∀x ∈ X ‖u(x)‖2 ≤ c

∫|x|2f dµ. Inspired

by this, Maurey developed in his thesis [98] an extensive factorization theory for operators eitherwith range Lp or with domain a subspace of Lp. Among his many results, he showed that for anysubspace X ⊂ Lp(µ) p > 2 and any operator u : E → Y into a Hilbert space Y (or a space of cotype2), there is a probability density f as above such that ∀x ∈ X ‖u(x)‖ ≤ C(p)‖u‖(

∫|x|2f1−2/pdµ)1/2

where the constant C(p) is independent of u.

32

Note that multiplication by f12− 1

p is an operator of norm ≤ 1 from Lp(µ) to L2(µ). For generalsubspaces X ⊂ Lp, the constant C(p) is unbounded when p → ∞ but if we restrict to X = Lp wefind a bounded constant C(p). In fact if we take Y = ℓ2 and let p → ∞ in that case we obtainroughly the little GT (see Theorem 5.2 above). It was thus natural to ask whether non-commutativeanalogues of these results hold. This question was answered first by the Khintchine inequality fornon-commutative Lp of [89]. Once this is known, the analogue of Maurey’s result follows by thesame argument as his, and the constant C(p) for this generalization remains of course unboundedwhen p → ∞. The case of a bounded operator u : Lp(τ) → Y on a non-commutative Lp-space(instead of a subspace) turned out to be much more delicate: The problem now was to obtain C(p)bounded when p → ∞, thus providing a generalization of the non-commutative little GT. Thisquestion was completely settled in [93] and in a more general setting in [95].

One of Rosenthal’s main achievements in [134] was the proof that for any reflexive subspaceE ⊂ L1(µ) the adjoint quotient mapping u∗ : L∞(µ) → E∗ was p′-summing for some p′ < ∞, andhence the inclusion E ⊂ L1(µ) factored through Lp(µ) in the form E → Lp(µ) → L1(µ) wherethe second arrow represents a multiplication operator by a nonnegative function in Lp′(µ). Thecase when E is a Hilbert space could be considered essentially as a corollary of the little GT, butreflexive subspaces definitely required deep new ideas.

Here again it was natural to investigate the non-commutative case. First in [114] the authorproved a non-commutative version of the so-called Nikishin–Maurey factorization through weak-Lp (this also improved the commutative factorization). However, this version was only partiallysatisfactory. A further result is given in [132], but a fully satisfactory non-commutative analogueof Rosenthal’s result was finally achieved by Junge and Parcet in [63]. These authors went onto study many related problems involving non-commutative Maurey factorizations, (also in theoperator space framework see [64]).

11. Best constants (Non-commutative case)

Let K ′G (resp. k′

G) denote the best possible constant in (7.3) (resp. (7.4)). Thus (7.3) and (7.4)show that K ′

G ≤ 2 and k′G ≤ 2. Note that, by the same simple argument as for (5.3), we have

k′G ≤ K ′

G.

We will show in this section (based on [47]) that, in sharp contrast with the commutative case, wehave

k′G = K ′

G = 2.

By the same reasoning as for Theorem 5.2 the best constant in (8.2) is equal to√

k′G, and hence√

2 is the best possible constant in (8.2).The next Lemma is the non-commutative version of Lemma 5.3.

Lemma 11.1. Let (M, τ) be a von Neumann algebra equipped with a normal, faithful, semi-finitetrace τ . Consider gj ∈ L1(τ) and xj ∈ M (1 ≤ j ≤ N) and positive numbers a, b. Assume that (gj)and (xj) are biorthogonal (i.e. 〈gi, xj〉 = τ(x∗

jgi) = 0 if i 6= j and = 1 otherwise) and such that

∀(αj) ∈ CN a(

∑|αj |2)1/2 ≥ ‖

∑αjgj‖L1(τ) and max

∥∥∥∑

x∗jxj

∥∥∥1/2

M,∥∥∥∑

xjx∗j

∥∥∥1/2

M

≤ b

√N.

Then b−2a−2 ≤ k′G (and a fortiori b−2a−2 ≤ K ′

G).

Proof. We simply repeat the argument used for Lemma 5.3.

33

Lemma 11.2 ([47]). Consider an integer n ≥ 1. Let N = 2n + 1 and d =(2n+1

n

)=

(2n+1n+1

). Let τd

denote the normalized trace on the space Md of d× d (complex) matrices. There are x1, · · · , xN inMd such that τd(x

∗i xj) = δij for all i, j (i.e. (xj) is orthonormal in L2(τd)), satisfying

∑x∗

jxj =∑

xjx∗j = NI and hence : max

∥∥∥∑

x∗jxj

∥∥∥1/2

,∥∥∥∑

xjx∗j

∥∥∥1/2

≤

√N,

and moreover such that

(11.1) ∀(αj) ∈ CN ‖

∑αjxj‖L1(τd) = ((n + 1)/(2n + 1))1/2(

∑|αj |2)1/2.

Proof. Let H be a (2n + 1)-dimensional Hilbert space with orthonormal basis e1, · · · , e2n+1. Wewill implicitly use analysis familiar in the context of the antisymmetric Fock space. Let H∧k denotethe antisymmetric k-fold tensor product. Note that dim(H∧k) = dim(H∧(2n+1−k)) =

(2n+1

k

)and

hence dim(H∧(n+1)) = dim(H∧n) = d. Let

cj : H∧n → H∧(n+1) and c∗j : H∧(n+1) → H∧n

be respectively the “creation” operator associated to ej (1 ≤ j ≤ 2n + 1) defined by cj(h) = ej ∧ hand its adjoint the so-called “annihilation” operator. For any ordered subset J ⊂ [1, · · · , N ], sayJ = j1, . . . , jk with j1 < · · · < jk we denote eJ = ej1 ∧ · · · ∧ ejk

. Recall that eJ | |J | = kis an orthonormal basis of H∧k. It is easy to check that cjc

∗j is the orthogonal projection from

H∧n+1 onto spaneJ | |J | = n + 1, j ∈ J, while c∗jcj is the orthogonal projection from H∧n onto

spaneJ | |J | = n, j 6∈ J. In addition, for each J with |J | = n + 1, we have∑2n+1

j=1 cjc∗j (eJ) =

|J |eJ = (n + 1)eJ , and similarly if |J | = n we have∑2n+1

j=1 c∗jcj(eJ) = (N − |J |)eJ = (n + 1)eJ .

(Note that∑2n+1

j=1 cjc∗j is the celebrated “number operator”). Therefore,

(11.2)∑2n+1

j=1cjc

∗j = (n + 1)IH∧(n+1) and

∑2n+1

j=1c∗jcj = (n + 1)IH∧(n) .

In particular this implies τd(∑2n+1

j=1 c∗jcj) = n + 1, and since, by symmetry τd(c∗jcj) = τd(c

∗1c1) for

all j, we must have τd(c∗1c1) = (2n + 1)−1(n + 1). Moreover, since c∗1c1 is a projection, we also

find τd(|c1|) = τd(|c1|2) = τd(c∗1c1) = (2n + 1)−1(n + 1) (which can also be checked by elementary

combinatorics).There is a well known “rotational invariance” in this context (related to what physicists call “secondquantization”), from which we may extract the following fact. Recall N = 2n + 1. Consider(αj) ∈ C

N and let h =∑

αjej ∈ H. Assume ‖h‖ = 1. Then roughly “creation by h” is equivalentto creation by e1, i.e. to c1. More precisely, Let U be any unitary operator on H such that Ue1 = h.Then U∧k is unitary on H∧k for any k, and it is easy to check that (U∧n)−1(

∑αjcj)U

∧(n+1) = c1.This implies

(11.3) τd(|∑

αjcj |2) = τd(|c1|2) and τd(|∑

αjcj |) = τd(|c1|).

Since dim(H∧n+1) = dim(H∧n) = d, we may (choosing orthonormal bases) identify cj with anelement of Md. Let then xj = cj(

2n+1n+1 )1/2. By the first equality in (11.3), (xj) are orthonormal in

L2(τd) and by (11.2) and the rest of (11.3) we have the announced result.

The next statement and (3.2) imply, somewhat surprisingly, that K ′G 6= KG.

Theorem 11.3 ([47]).k′

G = K ′G = 2.

34

Proof. By Lemma 11.2, the assumptions of Lemma 11.1 hold with a = 2−1/2 and b = 1. Thuswe have k′

G ≥ (ab)−2 = 2, and hence K ′G ≥ k′

G ≥ 2. But by Theorem 7.1 we already know thatK ′

G ≤ 2.

The next result is in sharp contrast with the commutative case: for the classical Khintchineinequality in L1 for the Steinhaus variables (sj), Sawa ([136]) showed that the best constant is1/‖gC

1 ‖1 = 2/√

π <√

2 (and of course this obviously holds as well for the sequence (gCj )).

Theorem 11.4 ([48]). The best constants in (9.5) and (9.7) are given by cC1 = cC

1 =√

2.

Proof. By Theorem 9.1 we already know the upper bounds, so it suffices to prove cC1 ≥

√2

and cC1 ≥

√2. The reasoning done to prove (5.7) can be repeated to show that k′

G ≤ cC1 and

k′G ≤ cC

1 , thus yielding the lower bounds. Since it is instructive to exhibit the culprits re-alizing the equality case, we include a direct deduction. By (9.3) and Lemma 11.2 we haveN ≤ |||(xj)|||1 max‖(xj)‖C , ‖(xj)‖R ≤ |||(xj)|||1N1/2, and hence N1/2 ≤ |||(xj)|||1. But by(11.1) we have

∫‖∑

sjxj‖L1(τd) = ((n + 1)/(2n + 1))1/2N1/2 so we must have N1/2 ≤ |||(xj)|||1 ≤cC1 ((n+1)/(2n+1))1/2N1/2 and hence we obtain ((2n+1)/(n+1))1/2 ≤ cC

1 , which, letting n → ∞,yields

√2 ≤ cC

1 . The proof that cC1 ≥

√2 is identical since by the strong law of large numbers

(∫

N−1∑N

1 |gCj |2)1/2 → 1 when n → ∞.

Remark 11.5. Similar calculations apply for Theorem 9.5: one finds that for any 1 ≤ p ≤ 2,cCp ≥ 21/p−1/2 and also cC

p ≥ 21/p−1/2. Note that these lower bounds are trivial for the sequence

(εj), e.g. the fact that c1 ≥√

2 is trivial already from the classical case (just consider ε1 + ε2 in(5.4)), but then in that case the proof of Theorem 9.1 from [48] only gives us c1 ≤

√3. So the best

value of c1 is still unknown !

12. C∗C∗

C∗-algebra tensor products, Nuclearity

The (maximal and minimal) tensor products of C∗-algebras were introduced in the late 1950’s (byGuichardet and Turumaru). The underlying idea most likely was inspired by Grothendieck’s workon the injective and projective Banach space tensor products.

Let A, B be two C∗-algebras. Note that their algebraic tensor product A ⊗ B is a ∗-algebra.For any t =

∑aj ⊗ bj ∈ A ⊗ B we define

‖t‖max = sup‖ϕ(t)‖B(H)

where the sup runs over all H and all ∗-homomorphisms ϕ : A ⊗ B → B(H). Equivalently, wehave

‖t‖max = sup

∥∥∥∑

σ(aj)ρ(bj)∥∥∥

B(H)

where the sup runs over all H and all possible pairs of ∗-homomorphisms σ : A → B(H), ρ : B →B(H) (over the same B(H)) with commuting ranges.

Since we have automatically ‖σ‖ ≤ 1 and ‖ρ‖ ≤ 1, we note that ‖t‖max ≤ ‖t‖∧. The completionof A⊗B equipped with ‖ · ‖max is denoted by A⊗max B and is called the maximal tensor product.This enjoys the following “projectivity” property: If I ⊂ A is a closed 2-sided (self-adjoint) ideal,then I ⊗max B ⊂ A ⊗max B (this is special to ideals) and the quotient C∗-algebra A/I satisfies

(12.1) (A/I) ⊗max (B) ≃ (A ⊗max B)/I ⊗max B.

35

The minimal tensor product can be defined as follows. Let σ : A → B(H) and ρ : B → B(K) beisometric ∗-homomorphisms. We set ‖t‖min = ‖(σ ⊗ ρ)(t)‖B(H⊗2K). It can be shown that this isindependent of the choice of such a pair (σ, ρ). We have

‖t‖min ≤ ‖t‖max.

Indeed, in the unital case just observe a → σ(a)⊗1 and b → 1⊗ρ(b) have commuting ranges. (Thegeneral case is similar using approximate units.)

The completion of (A ⊗ B, ‖ · ‖min) is denoted by A ⊗min B and is called the minimal tensorproduct. It enjoys the following “injectivity” property: Whenever A1 ⊂ A and B1 ⊂ B are C∗-subalgebras, we have

(12.2) A1 ⊗min B1 ⊂ A ⊗min B (isometric embedding).

However, in general ⊗min (resp. ⊗max) does not satisfy the “projectivity” (12.1) (resp. “injectivity”(12.2)).

By a C∗-norm on A ⊗ B, we mean any norm α adapted to the ∗-algebra structure and inaddition such that α(t∗t) = α(t)2. Equivalently this means that there is for some H a pair of∗-homomorphism σ : A → B(H), ρ : B → B(H) with commuting ranges such that ∀t =

∑aj ⊗ bj

α(t) =∥∥∥∑

σ(aj)ρ(bj)∥∥∥

B(H).

It is known that necessarily‖t‖min ≤ α(t) ≤ ‖t‖max.

Thus ‖ · ‖min and ‖ · ‖max are respectively the smallest and largest C∗-norms on A ⊗ B. Thecomparison with the Banach space counterparts is easy to check:

(12.3) ‖t‖∨ ≤ ‖t‖min ≤ ‖t‖max ≤ ‖t‖∧.

Remark. We recall that a C∗-algebra admits a unique C∗-norm. Therefore two C∗-norms that areequivalent on A ⊗ B must necessarily be identical (since the completion has a unique C∗-norm).

Definition 12.1. A pair of C∗-algebras (A, B) will be called a nuclear pair if A⊗min B = A⊗max Bi.e.

∀t ∈ A ⊗ B ‖t‖min = ‖t‖max.

Equivalently, this means that there is a unique C∗-norm on A ⊗ B.

Definition 12.2. A C∗-algebra A is called nuclear if for any C∗-algebra B, (A, B) is a nuclearpair.

The reader should note of course the analogy with Grothendieck’s previous notion of a “nuclearlocally convex space.” Note however that a Banach space E is nuclear in his sense (i.e. E⊗F = E⊗Ffor any Banach space) iff it is finite dimensional.

The following classical result is due independently to Choi-Effros and Kirchberg (see e.g. [18]).

Theorem 12.3. A unital C∗-algebra A is nuclear iff there is a net of finite rank maps of the formA

vα−→Mnα

wα−→A where vα, wα are maps with ‖vα‖cb ≤ 1, ‖wα‖cb ≤ 1, which tends pointwise to theidentity.

36

For example, all commutative C∗-algebras, the algebra K(H) of all compact operators on aHilbert space H, the Cuntz algebra or more generally all “approximately finite dimensional” C∗-algebras are nuclear.Whereas for counterexamples, let FN denote the free group on N generators, with 2 ≤ N ≤ ∞.(When N = ∞ we mean of course countably infinitely many generators.) The first non-nuclearC∗-algebras were found using free groups: both the full and reduced C∗-algebra of FN are notnuclear. Let us recall their definitions.For any discrete group G, the full (resp. reduced) C∗-algebra of G, denoted by C∗(G) (resp. C∗

λ(G))is defined as the C∗-algebra generated by the universal unitary representation of G (resp. the leftregular representation of G acting by translation on ℓ2(G)).

It turns out that “nuclear” is the analogue for C∗-algebras of “amenable” for groups. Indeed,from work of Lance (see [141]) it is known that C∗(G) or C∗

λ(G) is nuclear iff G is amenable(and in that case C∗(G) = C∗

λ(G)). More generally, an “abstract” C∗-algebra A is nuclear iff itis amenable as a Banach algebra (in B.E. Johnson’s sense). This means by definition that anybounded derivation D : A → X∗ into a dual A-module is inner. The fact that amenable impliesnuclear was proved by Connes as a byproduct of his deep investigation of injective von Neumannalgebras: A is nuclear iff the von Neumann algebra A∗∗ is injective, i.e. iff (assuming A∗∗ ⊂ B(H))there is a contractive projection P : B(H) → A∗∗, cf. Choi–Effros [21]. The converse implicationnuclear ⇒ amenable was proved by Haagerup in [44]. This crucially uses the non-commutative GT(in fact since nuclearity implies the approximation property, the original version of [112] is sufficienthere). For emphasis, let us state it:

Theorem 12.4. A C∗-algebra is nuclear iff it is amenable (as a Banach algebra).

Note that this implies that nuclearity passes to quotients (by a closed two sided ideal) but thatis not at all easy to see directly on the definition of nuclearity.

While the meaning of nuclearity for a C∗-algebra seems by now fairly well understood, it is notso for pairs, as reflected by Problem 12.9 below, proposed and emphasized by Kirchberg [77]. Inthis context, Kirchberg [77] proved the following striking result (a simpler proof was given in [119]):

Theorem 12.5. The pair (C∗(F∞), B(ℓ2)) is a nuclear pair.

Note that any separable C∗-algebra is a quotient of C∗(F∞), so C∗(F∞) can be viewed as“projectively universal.” Similarly, B(ℓ2) is “injectively universal” in the sense that any separableC∗-algebra embeds into B(ℓ2).

Let Aop denote the “opposite” C∗-algebra of A, i.e. the same as A but with the product in reverseorder. For various reasons, it was conjectured by Lance that A is nuclear iff the pair (A, Aop) isnuclear. This conjecture was supported by the case A = C∗

λ(G) and many other special cases whereit did hold. Nevertheless a remarkable counterexample was found by Kirchberg in [77]. (See [115]for the Banach analogue of this counterexample, an infinite dimensional (i.e. non-nuclear!) Banachspace X such that the injective and projective tensor norms are equivalent on X ⊗ X.) Kirchbergthen wondered whether one could simply take either A = B(H) (this was disproved in [66] seeRemark 12.8 below) or A = C∗(F∞) (this is the still open Problem 12.9 below). Note that if eitherA = B(H) or A = C∗(F∞)), we have A ≃ Aop, so the pair (A, Aop) can be replaced by (A, A).Note also that, for that matter, F∞ can be replaced by F2 or any non Abelian free group (sinceF∞ ⊂ F2).

For our exposition, it will be convenient to adopt the following definitions (equivalent to themore standard ones by [77]).

37

Definition 12.6. Let A be a C∗-algebra. We say that A is WEP if (A, C∗(F∞)) is a nuclear pair.We say that A is LLP if (A, B(ℓ2)) is a nuclear pair. We say that A is QWEP if it is a quotient(by a closed, self-adjoint, 2-sided ideal) of a WEP C∗-algebra.

Here WEP stands for weak expectation property (and LLP for local lifting property). AssumingA∗∗ ⊂ B(H), A is WEP iff there is a completely positive T : B(H) → A∗∗ (“a weak expectation”)such that T (a) = a for all a in A.

Kirchberg actually proved the following generalization of Theorem 12.5:

Theorem 12.7. If A is LLP and B is WEP then (A, B) is a nuclear pair.

Remark 12.8. S. Wasserman proved in [152] that B(H) and actually any von Neumann algebra M(except for a very special class) are not nuclear. The exceptional class is formed of all finite directsums of tensor products of a commutative algebra with a matrix algebra (“finite type I”). Theproof in [66] that (B(H), B(H)) (or (M, M)) is not a nuclear pair is based on Kirchberg’s idea that,otherwise, the set formed of all the finite dimensional operator spaces, equipped with the distanceδcb(E, F ) = log dcb(E, F ) (dcb is defined in (13.2) below), would have to be a separable metricspace. The original proof in [66] that the latter metric space (already for 3-dimensional operatorspaces) is actually non-separable used GT for exact operator spaces (described in §17 below), butsoon after, several different, more efficient proofs were found (see chapters 21 and 22 in [123]).

The next problem is currently perhaps the most important open question in Operator AlgebraTheory. Indeed, as Kirchberg proved (see the end of [77]), this is also equivalent to a fundamentalquestion raised by Alain Connes [22] on von Neumann algebras: whether any separable II1-factor(or more generally any separable finite von Neumann algebra) embeds into an ultraproduct ofmatrix algebras. We refer the reader to [104] for an excellent exposition of this topic.

Problem 12.9. Let A = C∗(FN ) (N > 1). Is there a unique C∗-norm on A ⊗ A? Equivalently, is(A, A) a nuclear pair?

Equivalently:

Problem 12.10. Is every C∗-algebra QWEP ?

Curiously, there seems to be a connection between GT and Problem 12.9. Indeed, like GT, thelatter problem can be phrased as the identity of two tensor products on ℓ1⊗ℓ1, or more precisely onMn ⊗ ℓ1 ⊗ ℓ1, which might be related to the operator space version of GT presented in the sequel.To explain this, we first reformulate Problem 12.9.

Let A = C∗(F∞). Let (Uj)j≥1 denote the unitaries of C∗(F∞) that correspond to the freegenerators of A. For convenience of notation we set U0 = I. It is easy to check that the closedlinear span E ⊂ A of Uj | j ≥ 0 in A is isometric to ℓ1 and that E generates A as a C∗-algebra.

Fix n ≥ 1. Consider a finitely supported family a = ajk | j, k ≥ 0 with ajk ∈ Mn for allj, k ≥ 0. We denote, again for A = C∗(F∞):

‖a‖min =∥∥∥∑

ajk ⊗ Uj ⊗ Uk

∥∥∥Mn(A⊗minA)

‖a‖max =∥∥∥∑

ajk ⊗ Uj ⊗ Uk

∥∥∥Mn(A⊗maxA)

.

Then, on one hand, going back to the definitions, it is easy to check that

(12.4) ‖a‖max = sup

∥∥∥∑

ajk ⊗ ujvk

∥∥∥Mn(B(H))

38

where the supremum runs over all H and all possible unitaries uj , vk on the same Hilbert space Hsuch that ujvk = vkuj for all j, k. On the other hand, using the known fact that A embeds into adirect sum of matrix algebras (due to M.D. Choi, see e.g. [18, §7.4]), one can check that

(12.5) ‖a‖min = sup

∥∥∥∑

ajk ⊗ ujvk

∥∥∥Mn(B(H))

where the sup is as in (12.4) except that we restrict it to all finite dimensional Hilbert spaces H.We may ignore the restriction u0 = v0 = 1 because we can always replace (uj , vj) by (u−1

0 uj , vjv−10 )

without changing either (12.4) or (12.5).The following is implicitly in [119].

Proposition 12.11. Let A = C∗(F∞). The following assertions are equivalent:

(i) A ⊗min A = A ⊗max A (i.e. Problem 12.9 has a positive solution).

(ii) For any n ≥ 1 and any ajk | j, k ≥ 0 ⊂ Mn as above the norms (12.4) and (12.5) coincidei.e. ‖a‖min = ‖a‖max.

(iii) The identity ‖a‖min = ‖a‖max holds for all n ≥ 1 but merely for all families ajk in Mn

supported in the union of 0 × [0, 1, 2] and [0, 1, 2] × 0.Theorem 12.12 ([146]). Assume n = 1 and ajk ∈ R for all j, k. Then

‖a‖max = ‖a‖min,

and these norms coincide with the H ′-norm of∑

ajkej ⊗ ek in ℓ1 ⊗ ℓ1, that we denote by ‖a‖H′.Moreover, these are all equal to

(12.6) sup ‖∑

ajkujvk‖

where the sup runs over all N and all self-adjoint unitary N × N matrices such that ujvk = vkuj

for all j, k.

Proof. Recall (see (1.13)) that ‖a‖H′ ≤ 1 iff for any unit vectors xj , yk in a Hilbert space we have∣∣∣∑

ajk〈xj , yk〉∣∣∣ ≤ 1.

Note that, since ajk ∈ R, whether we work with real or complex Hilbert spaces does not affect thiscondition. The resulting H ′-norm is the same. We have trivially

‖a‖min ≤ ‖a‖max ≤ ‖a‖H′ ,

so it suffices to check ‖a‖H′ ≤ ‖a‖min. Consider unit vectors xj , yk in a real Hilbert space H.We may assume that ajk is supported in [1, . . . , n] × [1, . . . , n] and that dim(H) = n. Fromclassical facts on “spin systems” (Pauli matrices, Clifford algebras and so on), we know that thereare self-adjoint unitary matrices Xj , Yk such that XjYk = −YkXj for all j, k and a (vector) stateF on H such that F (XjYk) = i〈xj , yk〉 ∈ iR. Let Q = ( 0 1

1 0 ), P =(

0 i−i 0

). Note QP = −PQ and

(QP )11 = −i. Therefore if we set uj = Xj ⊗ Q, vk = Yk ⊗ P , and f = F ⊗ e11, we find self-adjointunitaries such that ujvk = vkuj for all j, k and f(ujvk) = 〈xj , yk〉 ∈ R. Therefore we conclude

∣∣∣∑

ajk〈xj , yk〉∣∣∣ =

∣∣∣f(∑

ajkujvk

)∣∣∣ ≤ ‖a‖min,

and hence ‖a‖H′ ≤ ‖a‖min. This proves ‖a‖H′ = ‖a‖min but also ‖∑ajkej ⊗ ek‖H′ ≤ (12.6). Since,

by (12.5), we have (12.6) ≤ ‖a‖min, the proof is complete.

39

The preceding equality ‖a‖max = ‖a‖min seems open when ajk ∈ C (i.e. even the case n = 1 inProposition 12.11 (ii) is open). However, we have

Proposition 12.13. Assume n = 1 and ajk ∈ C for all j, k ≥ 0 then ‖a‖max ≤ KCG‖a‖min.

Proof. By (12.3) we have

sup

∣∣∣∑

ajksjtk

∣∣∣∣∣∣∣ sj , tk ∈ C |sj | = |tk| = 1

= ‖a‖∨ ≤ ‖a‖min.

By Theorem 2.4 we have for any unit vectors x, y in H∣∣∣∑

ajk〈ujvkx, y〉∣∣∣ =

∣∣∣∑

ajk〈vkx, u∗jy〉

∣∣∣ ≤ KCG‖a‖∨ ≤ KC

G‖a‖min.

Actually this holds even without the assumption that the uj’s commute with the vk’s.

Remark. Although we are not aware of a proof, we believe that the equalities ‖a‖H′ = ‖a‖min and‖a‖H′ = ‖a‖max in Theorem 12.12 do not extend to the complex case. However, if [aij ] is a 2 × 2matrix, they do extend because, by [23, 144] for 2 × 2 matrices in the complex case (2.5) happensto be valid with K = 1 (while in the real case, for 2 × 2 matrices, the best constant is

√2).

See [128, 129, 28, 61, 32] for related contributions.

13. Operator spaces, c.b. maps, Minimal tensor product

The theory of operator spaces is rather recent. It is customary to date its birth with the 1987 thesisof Z.J. Ruan. This started a broad series of foundational investigations mainly by Effros-Ruan andBlecher-Paulsen. See [30, 123]. We will try to merely summarize the developments that relate toour main theme, meaning GT.

We start by recalling a few basic facts. First, a general unital C∗-algebra can be viewed asthe non-commutative analogue of the space C(Ω) of continuous functions on a compact set Ω.Secondly, any Banach space B can be viewed isometrically as a closed subspace of C(Ω): just takefor Ω the closed unit ball of B∗ and consider the isometric embedding taking x ∈ B to the functionω → ω(x).

Definition 13.1. An operator space E is a closed subspace E ⊂ A of a general (unital if we wish)C∗-algebra.

With the preceding two facts in mind, operator spaces appear naturally as “non-commutativeBanach spaces.” But actually the novelty in Operator space theory in not so much in the “spaces”as it is in the morphisms. Indeed, if E1 ⊂ A1, E2 ⊂ A2 are operator spaces, the natural morphismsu : E1 → E2 are the “completely bounded” linear maps that are defined as follows. First note thatif A is a C∗-algebra, the algebra Mn(A) of n × n matrices with entries in A is itself a C∗-algebraand hence admits a specific (unique) C∗-algebra norm. The latter norm can be described a bitmore concretely when one realizes (by Gelfand–Naimark) A as a closed self-adjoint subalgebra ofB(H) with norm induced by that of B(H). In that case, if [aij ] ∈ Mn(A) then the matrix [aij ]can be viewed as a single operator acting naturally on H ⊕ · · · ⊕ H (n times) and its norm isprecisely the C∗-norm of Mn(A). In particular the latter norm is independent of the embedding(or “realization”) A ⊂ B(H).

As a consequence, if E ⊂ A is any closed subspace, then the space Mn(E) of n × n matriceswith entries in E inherits the norm induced by Mn(A). Thus, we can associate to an operator spaceE ⊂ A, the sequence of Banach spaces Mn(E) | n ≥ 1.

40

Definition 13.2. A linear map u : E1 → E2 is called completely bounded (c.b. in short) if

(13.1) ‖u‖cbdef= sup

n≥1‖un : Mn(E1) → Mn(E2)‖ < ∞

where for each n ≥ 1, un is defined by un([aij ]) = [u(aij)]. One denotes by CB(E1, E2) the spaceof all such maps.

The following factorization theorem (proved by Wittstock, Haagerup and Paulsen independentlyin the early 1980’s) is crucial for the theory. Its origin goes back to major works by Stinespring(1955) and Arveson (1969) on completely positive maps.

Theorem 13.3. Consider u : E1 → E2. Assume E1 ⊂ B(H1) and E2 ⊂ B(H2). Then ‖u‖cb ≤ 1iff there is a Hilbert space H, a representation π : B(H1) → B(H) and operators V, W : H2 → Hwith ‖V ‖‖W‖ ≤ 1 such that

∀x ∈ E1 u(x) = V ∗π(x)W.

We refer the reader to [30, 123, 106] for more background.We say that u is a complete isomorphism (resp. complete isometry) if u is invertible and u−1

is c.b. (resp. if un is isometric for all n ≥ 1). We say that u : E1 → E2 is a completely isomorphicembedding if u induces a complete isomorphism between E1 and u(E1).

The following non-commutative analogue of the Banach–Mazur distance has proved quite useful,especially to compare finite dimensional operator spaces. When E, F are completely isomorphic weset

(13.2) dcb(E, F ) = inf‖u‖cb‖u−1‖cb

where the infimum runs over all possible isomorphisms u : E → F . We also set dcb(E, F ) = ∞ ifE, F are not completely isomorphic.Fundamental examples: Let us denote by eij the matrix units in B(ℓ2) (or in Mn). Let

C = span[ei1 | i ≥ 1] (“column space”) and R = span[e1j | j ≥ 1] (“row space”).

We also define the n-dimensional analogues:

Cn = span[ei1 | 1 ≤ i ≤ n] and Rn = span[e1j | 1 ≤ j ≤ n].

Then C ⊂ B(ℓ2) and R ⊂ B(ℓ2) are very simple but fundamental examples of operator spaces.Note that R ≃ ℓ2 ≃ C as Banach spaces but R and C are not completely isomorphic. Actuallythey are “extremely” non-completely isomorphic: one can even show that

n = dcb(Cn, Rn) = supdcb(E, F ) | dim(E) = dim(F ) = n.

Remark 13.4. It can be shown that the map on Mn that takes a matrix to its transposed hascb-norm = n (although it is isometric). In the opposite direction, the norm of a Schur multiplier onB(ℓ2) is equal to its cb-norm. As noticed early on by Haagerup (unpublished), this follows easilyfrom Proposition 2.9.

41

When working with an operator space E we rarely use a specific embedding E ⊂ A, however, wecrucially use the spaces CB(E, F ) for different operator spaces F . By (13.1) the latter are entirelydetermined by the knowledge of the sequence of normed (Banach) spaces

(13.3) Mn(E) | n ≥ 1.

This is analogous to the fact that knowledge of a normed (or Banach space before completion)boils down to that of a vector space equipped with a norm. In other words, we view the sequenceof norms in (13.3) as the analogue of the norm. Note however, that not any sequence of normson Mn(E) “comes from” a “realization” of E as operator space (i.e. an embedding E ⊂ A). Thesequences of norms that do so have been characterized by Ruan:

Theorem 13.5 (Ruan’s Theorem). Let E be a vector space. Consider for each n a norm αn onthe vector space Mn(E). Then the sequence of norms (αn) comes from a linear embedding of Einto a C∗-algebra iff the following two properties hold:

• ∀n, ∀x ∈ Mn(E) ∀a, b ∈ Mn αn(a.x.b) ≤ ‖a‖Mnαn(x) ‖b‖Mn

.

• ∀n, m ∀x ∈ Mn(E) ∀y ∈ Mm(E) αn+m(x⊕ y) = maxαn(x), αm(y), where we denote by

x ⊕ y the (n + m) × (n + m) matrix defined by x ⊕ y =

(x 00 y

).

Using this, several important constructions involving operator spaces can be achieved, althoughthey do not make sense a priori when one considers concrete embeddings E ⊂ A. Here is a list ofthe main such constructions.

Theorem 13.6. Let E be an operator space.

(i) Then there is a C∗-algebra B and an isometric embedding E∗ ⊂ B such that for any n ≥ 1

(13.4) Mn(E∗) ≃ CB(E, Mn) isometrically.

(ii) Let F ⊂ E be a closed subspace. There is a C∗-algebra B and an isometric embeddingE/F ⊂ B such that for all n ≥ 1

(13.5) Mn(E/F ) = Mn(E)/Mn(F ).

(iii) Let (E0, E1) be a pair of operator spaces, assumed compatible for the purpose of interpolation(cf. [9, 123]). Then for each 0 < θ < 1 there is a C∗-algebra Bθ and an isometric embedding(E0, E1)θ ⊂ Bθ such that for all n ≥ 1

(13.6) Mn((E0, E1)θ) = (Mn(E0), Mn(E1))θ.

In each of the identities (13.4), (13.5), (13.6) the right-hand side makes natural sense, thecontent of the theorem is that in each case the sequence of norms on the right hand side “comesfrom” a concrete operator space structure on E∗ in (i), on E/F in (ii) and on (E0, E1)θ in (iii).The proof reduces to the verification that Ruan’s criterion applies in each case.

The general properties of c.b. maps, subspaces, quotients and duals mimic the analogousproperties for Banach spaces, e.g. if u ∈ CB(E, F ), v ∈ CB(F, G) then vu ∈ CB(E, G) and

(13.7) ‖vu‖cb ≤ ‖v‖cb‖u‖cb;

42

also ‖u∗‖cb = ‖u‖cb, and if F is a closed subspace of E, F ∗ = E∗/F⊥, (E/F )∗ = F⊥ and E ⊂ E∗∗

completely isometrically.It is natural to revise our terminology slightly: by an operator space structure (o.s.s. in short)

on a vector (or Banach) space E we mean the data of the sequence of the norms in (13.3). We then“identify” the o.s.s. associated to two distinct embeddings, say E ⊂ A and E ⊂ B, if they lead toidentical norms in (13.3). (More rigorously, an o.s.s. on E is an equivalence class of embeddings asin Definition 13.1 with respect to the preceding equivalence.)

Thus, Theorem 13.6 (i) allows us to introduce a duality for operator spaces: E∗ equipped withthe o.s.s. defined in (13.4) is called the o.s. dual of E. Similarly E/F and (E0, E1)θ can now beviewed as operator spaces equipped with their respective o.s.s. (13.5) and (13.6).

The minimal tensor product of C∗-algebras induces naturally a tensor product for operatorspaces: given E1 ⊂ A1 and E2 ⊂ A2, we have E1⊗E2 ⊂ A⊗min A2 so the minimal C∗-norm inducesa norm on E1 ⊗ E2 that we still denote by ‖ · ‖min and we denote by E1 ⊗min E2 the completion.Thus E1 ⊗min E2 ⊂ A1 ⊗min A2 is an operator space.

Remark 13.7. It is useful to point out that the norm in E1 ⊗min E2 can be obtained from the o.s.s.of E1 and E2 as follows. One observes that any C∗-algebra can be completely isometrically (butnot as subalgebra) embedded into a direct sum

⊕i∈I

Mn(i) of matrix algebras for a suitable family of

integers n(i) | i ∈ I. Therefore using the o.s.s. of E1 we may assume

E1 ⊂⊕

i∈IMn(i)

and then we findE1 ⊗min E2 ⊂

⊕i∈I

Mn(i)(E2).

We also note the canonical (flip) identification:

(13.8) E1 ⊗min E2 ≃ E2 ⊗min E1.

Remark 13.8. In particular, taking E2 = E∗ and E1 = F we find using (13.4)

F ⊗min E∗ ⊂⊕

i∈IMn(i)(E

∗) =⊕

i∈I

CB(E, Mn(i)) ⊂ CB

(E,

⊕

i∈I

Mn(i)

)

and hence (using (13.8))

(13.9) E∗ ⊗min F ≃ F ⊗min E∗ ⊂ CB(E, F ) isometrically.

More generally, by Ruan’s theorem, the space CB(E, F ) can be given an o.s.s. by declaringthat Mn(CB(E, F )) = CB(E, Mn(F )) isometrically. Then (13.9) becomes a completely isometricembedding.

As already mentioned, ∀u : E → F , the adjoint u∗ : F ∗ → E∗ satisfies

(13.10) ‖u∗‖cb = ‖u‖cb.

Collecting these observations, we obtain:

Proposition 13.9. Let E, F be operator spaces and let C > 0 be a constant. The followingproperties of a linear map u : E → F ∗ are equivalent.

(i) ‖u‖cb ≤ C.

43

(ii) For any integers n, m the bilinear map Φn,m : Mn(E)×Mm(F ) → Mn ⊗Mm ≃ Mnm definedby

Φn,m([aij ], [bkℓ]) = [〈u(aij), bkℓ〉]satisfies ‖Φn,m‖ ≤ C.Explicitly, for any finite sums

∑r ar ⊗ xr ∈ Mn(E), and

∑s bs ⊗ ys ∈ Mm(F ) and for any

n × m scalar matrices α, β, we have

(13.11)

∣∣∣∣∣∑

r,s

tr(arαtbsβ

∗)〈uxr, ys〉∣∣∣∣∣ ≤ C‖α‖2‖β‖2

∥∥∥∑

ar ⊗ xr

∥∥∥Mn(E)

∥∥∥∑

bs ⊗ ys

∥∥∥Mm(F )

where ‖ · ‖2 denotes the Hilbert–Schmidt norm.

(iii) For any pair of C∗-algebras A, B, the bilinear map

ΦA,B : A ⊗min E × B ⊗min F −→ A ⊗min B

defined by ΦA,B(a ⊗ e, b ⊗ f) = a ⊗ b 〈ue, f〉 satisfies ‖ΦA,B‖ ≤ C.Explicitly, whenever

∑n1 ai ⊗ xi ∈ A ⊗ E,

∑m1 bj ⊗ yj ∈ B ⊗ F , we have

∥∥∥∥∥∥

∑

i,j

ai ⊗ bj〈u(xi), yj〉

∥∥∥∥∥∥A⊗minB

≤ C∥∥∥∑

ai ⊗ xi

∥∥∥min

∥∥∥∑

bj ⊗ yj

∥∥∥min

.

Let H1, H2 be Hilbert spaces. Clearly, the space B(H1, H2) can be given a natural o.s.s.: onejust considers a Hilbert space H such that H1 ⊂ H and H2 ⊂ H (for instance H = H1 ⊕ H2)so that we have H ≃ H1 ⊕ K1, H ≃ H2 ⊕ K2 and then we embed B(H1, H2) into B(H) via themapping represented matricially by

x 7→(

0 0x 0

).

It is easy to check that the resulting o.s.s. does not depend on H or on the choice of embeddingsH1 ⊂ H, H2 ⊂ H.

In particular, using this we obtain o.s.s. on B(C, H) and B(H∗, C) for any Hilbert space H. Wewill denote these operator spaces as follows

(13.12) Hc = B(C, H) Hr = B(H∗, C).

The spaces Hc and Hr are isometric to H as Banach spaces, but are quite different as o.s. WhenH = ℓ2 (resp. H = ℓn

2 ) we recover the row and column spaces, i.e. we have

(ℓ2)c = C, (ℓ2)r = R, (ℓn2 )c = Cn, (ℓn

2 )r = Rn.

The next statement incorporates observations made early on by the founders of operator spacetheory, namely Blecher-Paulsen and Effros-Ruan. We refer to [30, 123] for more references.

Theorem 13.10. Let H, K be arbitrary Hilbert spaces. Consider a linear map u : E → F .

(i) Assume either E = Hc and F = Kc or E = Hr and F = Kr then

CB(E, F ) = B(E, F ) and ‖u‖cb = ‖u‖.

44

(ii) If E = Rn (resp. E = Cn) and if ue1j = xj (resp. uej1 = xj), then

‖u‖cb = ‖(xj)‖C (resp. ‖u‖cb = ‖(xj)‖R).

(iii) Assume either E = Hc and F = Kr or E = Hr and F = Kc. Then u is c.b. iff it isHilbert–Schmidt and, denoting by ‖ · ‖2 the Hilbert–Schmidt norm, we have

‖u‖cb = ‖u‖2.

(iv) Assume F = Hc. Then ‖u‖cb ≤ 1 iff for all finite sequences (xj) in E we have

(13.13)(∑

‖uxj‖2)1/2

≤ ‖(xj)‖C .

(v) Assume F = Kr. Then ‖u‖cb ≤ 1 iff for all finite sequences (xj) in E we have

(∑‖uxj‖2

)1/2≤ ‖(xj)‖R.

Proof of (i) and (iv): (i) Assume say E = F = Hc. Assume u ∈ B(E, F ) = B(H). The mappingx → ux is identical to the mapping B(C, H) → B(C, H) of left multiplication by u. The latter isclearly cb with c.b. norm at most ‖u‖. Therefore ‖u‖cb ≤ ‖u‖, and the converse is trivial.(iii) Assume ‖u‖cb ≤ 1. Note that

‖(xj)‖C = ‖v : Rn → E‖cb

where v is defined by vej1 = xj . (This follows from the identity R∗n ⊗min E = CB(Rn, E) and

R∗n = Cn.) Therefore we have

‖uv : Rn → F‖cb ≤ ‖u‖cb‖(xj)‖C

and hence by (ii)

‖uv‖2 =(∑

j‖uvej1‖2

)1/2=

(∑‖xj‖2

)1/2≤ ‖u‖cb‖(xj)‖C .

This proves the only if part.To prove the converse, assume E ⊂ A (A being a C∗-algebra). Note ‖(xj)‖2

C = sup∑ f(x∗jxj) |

f state on A. Then by Proposition 4.5, (13.13) implies that there is a state f on A such that

(13.14) ∀x ∈ E ‖ux‖2 ≤ f(x∗x) = ‖πf (x)ξf‖2

where πf denotes the representation on A associated to f on a Hilbert space Hf and ξf is acyclic unit vector in Hf . By (13.14) there is an operator b : Hf → K with ‖b‖ ≤ 1 such thatux = bπf (x)ξf . But then this is precisely the canonical factorization of c.b. maps described inTheorem 13.3, but here for the map x 7→ ux ∈ B(C, K) = Kc. Thus we conclude ‖u‖cb ≤ 1.

Remark 13.11. Assume E ⊂ A (A being a C∗-algebra). By Proposition 4.5, (iv) (resp. (v)) holdsiff there is a state f on A such that

∀x ∈ E ‖ux‖2 ≤ f(x∗x) (resp. ∀x ∈ E ‖ux‖2 ≤ f(xx∗)).

45

Remark 13.12. The operators u : A → ℓ2 (A being a C∗-algebra) such that for some constant cthere is a state f on A such that

∀x ∈ A ‖ux‖2 ≤ c(f(x∗x)f(xx∗))1/2,

have been characterized in [126] as those that are completely bounded from A (or from an exactsubspace E ⊂ A) to the operator Hilbert space OH. See [124] for more on this.

Remark 13.13. As Banach spaces, we have c∗0 = ℓ1, ℓ∗1 = ℓ∞ and more generally L∗1 = L∞.

Since the spaces c0 and L∞ are C∗-algebras, they admit a natural specific o.s.s. (determinedby the unique C∗-norms on Mn ⊗ c0 and Mn ⊗ L∞), therefore, by duality we may equip alsoℓ1 (or L1 ⊂ L∗

∞) with a natural specific o.s.s. called the “maximal o.s.s.”. In the case of ℓ1,it is easy to describe: Indeed, it is a simple exercise (recommended to the reader) to check thatthe embedding ℓ1 ≃ spanUj ⊂ C∗(F∞), already considered in §sec10 constitutes a completelyisometric realization of this operator space ℓ1 = c∗0 (recall (Uj)j≥1 denote the unitaries of C∗(F∞)that correspond to the free generators) .

14. Haagerup tensor product

Surprisingly, Operator spaces admit a special tensor product, the Haagerup tensor product, thatdoes not really have any counterpart for Banach spaces. Based on unpublished work of Haageruprelated to GT, in 1987 Effros and Kishimoto popularized it under this name. At first only its normwas considered, but somewhat later, its o.s.s. emerged as a crucial concept to understand completelypositive and c.b. multilinear maps, notably in fundamental work by Christensen and Sinclair,continued by many authors, namely Paulsen, Smith, Blecher, Effros and Ruan. See [30, 123] forprecise references.

We now formally introduce the Haagerup tensor product E ⊗h F of two operator spaces.Assume E ⊂ A, F ⊂ B. Consider a finite sum t =

∑xj ⊗ yj ∈ E ⊗ F . We define

(14.1) ‖t‖h = inf‖(xj)‖R‖(yj)‖Cwhere the infimum runs over all possible representations of t. More generally, given a matrixt = [tij ] ∈ Mn(E ⊗ F ) we consider factorizations of the form

tij =N∑

k=1

xik ⊗ ykj

with x = [xik] ∈ Mn,N (E), y = [ykj ] ∈ MN,n(F ) and we define

(14.2) ‖t‖Mn(E⊗hF ) = inf‖x‖Mn,N (E)‖y‖MN,n(F )the inf being over all N and all possible such factorizations.

It turns out that this defines an o.s.s. on E ⊗h F , so there is a C∗-algebra C and an embeddingE⊗h F ⊂ C that produces the same norms as in (14.2). This can be deduced from Ruan’s theorem,but one can also take C = A∗B (full free product of C∗-algebras) and use the “concrete” embedding

∑xj ⊗ yj −→

∑xj · yj ∈ A ∗ B.

Then this is completely isometric. See e.g. [123, §5].In particular, since we have an (automatically completely contractive) ∗-homomorphism A ∗ B →A ⊗min B, it follows

(14.3) ‖t‖Mn(E⊗minF ) ≤ ‖t‖Mn(E⊗hF ).

46

For any linear map u : E → F between operator spaces we denote by γr(u) (resp. γc(u)) theconstant of factorization of u through a space of the form Hr (resp. Kc), as defined in (13.12).More precisely, we set

(14.4) γr(u) = inf‖u1‖cb‖u2‖cb (resp. γc(u) = inf‖u1‖cb‖u2‖cb)

where the infimum runs over all possible Hilbert spaces H, K and all factorizations

Eu1−→ Z

u2−→ F

of u through Z with Z = Hr (resp. Z = Kc). (See also [102] for a symmetrized version of theHaagerup tensor product, for maps factoring through Hr ⊕ Kc).

Theorem 14.1. Let E ⊂ A, F ⊂ B be operator spaces (A, B being C∗-algebras). Consider a linearmap u : E → F ∗ and let ϕ : E × F → C be the associated bilinear form. Then ‖ϕ‖(E⊗hF )∗ ≤ 1 iffeach of the following equivalent conditions holds:

(i) For any finite sets (xj), (yj) in E and F respectively we have

∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ ‖(xj)‖R‖(yj)‖C .

(ii) There are states f, g respectively on A and B such that

∀(x, y) ∈ E × F |〈ux, y〉| ≤ f(xx∗)1/2g(y∗y)1/2.

(iii) γr(u) ≤ 1.

Moreover if u has finite rank, i.e. ϕ ∈ E∗ ⊗ F ∗ then

(14.5) ‖ϕ‖(E⊗hF )∗ = ‖ϕ‖E∗⊗hF ∗ .

Proof. (i) ⇔ ‖ϕ‖(E⊗hF )∗ ≤ 1 is obvious. (i) ⇔ (ii) follows by the Hahn–Banach type argument(see §4). (ii) ⇔ (iii) follows from Remark 13.11.

Remark 14.2. If we replace A, B by the “opposite” C∗-algebras (i.e. we reverse the product), weobtain induced o.s.s. on E and F denoted by Eop and F op. Another more concrete way to lookat this, assuming E ⊂ B(ℓ2), is that Eop consists of the transposed matrices of those in E. It iseasy to see that Rop = C and Cop = R. Applying Theorem 14.1 to Eop ⊗h F op, we obtain theequivalence of the following assertions:

(i) For any finite sets (xj), (yj) in E and F

∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ ‖(xj)‖C‖(yj)‖R.

(ii) There are states f, g on A, B such that

∀(x, y) ∈ E × F |〈ux, y〉| ≤ f(x∗x)1/2g(yy∗)1/2

(iii) γc(u) ≤ 1.

47

Remark 14.3. Since we have trivially ‖u‖cb ≤ γr(u) and ‖u‖cb ≤ γc(u), the three equivalent prop-erties in Theorem 14.1 and also the three appearing in the preceding Remark 14.2 each imply that‖u‖cb ≤ 1.

Remark 14.4. The most striking property of the Haagerup tensor product is probably its selfduality.There is no Banach space analogue for this. By “selfduality” we mean that if either E or F is finitedimensional, then (E ⊗h F )∗ = E∗ ⊗h F ∗ completely isometrically, i.e. these coincide as operatorspaces and not only (as expressed by (14.5)) as Banach spaces.

Remark 14.5. Returning to the notation in (1.12), let E and F be commutative C∗-algebras (e.g.E = F = ℓn

∞) with their natural o.s.s.. Then, a simple verification from the definitions shows thatfor any T ∈ E ⊗ F we have on one hand ‖T‖min = ‖T‖∨ and on the other hand ‖T‖h = ‖T‖H .Moreover, if u : E∗ → F is the associated linear map, we have ‖u‖cb = ‖u‖ and γr(u) = γc(u) =γ2(u). Using the selfduality described in the preceding remark, we find that for any T ′ ∈ L1 ⊗ L1

(e.g. T ′ ∈ ℓn1 ⊗ ℓn

1 ) we have ‖T ′‖h = ‖T ′‖H′ .

15. The operator Hilbert space OH

We call an operator space Hilbertian if the underlying Banach space is isometric to a Hilbert space.For instance R, C are Hilbertian, but there are also many more examples in C∗-algebra relatedtheoretical physics, e.g. Fermionic or Clifford generators of the so-called CAR algebra (CAR standsfor canonical anticommutation relations), generators of the Cuntz algebras, free semi-circular orcircular systems in Voiculescu’s free probability theory, . . .. None of them however is self-dual.Thus, the next result came somewhat as a surprise. (Notation: if E is an operator space, sayE ⊂ B(H), then E is the complex conjugate of E equipped with the o.s. structure correspondingto the embedding E ⊂ B(H) = B(H).)

Theorem 15.1 ([120]). Let H be an arbitrary Hilbert space. There exists, for a suitable H, aHilbertian operator space EH ⊂ B(H) isometric to H such that the canonical identification (derivedfrom the scalar product) E∗

H → EH is completely isometric. Moreover, the space EH is unique upto complete isometry. Let (Ti)i∈I be an orthonormal basis in EH . Then, for any n and any finitely

supported family (ai)i∈I in Mn, we have ‖∑ai ⊗ Ti‖Mn(EH) = ‖∑

ai ⊗ ai‖1/2M

n2.

When H = ℓ2, we denote the space EH by OH and we call it the “operator Hilbert space”.Similarly, we denote it by OHn when H = ℓn

2 and by OH(I) when H = ℓ2(I).The norm of factorization through OH, denoted by γoh, can be studied in analogy with (1.10),

and the dual norm is identified in [120] in analogy with Corollary 4.3.Although this is clearly the “right” operator space analogue of Hilbert space, we will see shortly

that, in the context of GT, it is upstaged by the space R ⊕ C.The space OH has rather striking complex interpolation properties. For instance, we have a

completely isometric identity (R, C) 12≃ OH, where the pair (R, C) is viewed as “compatible” using

the transposition map x → tx from R to C which allows to view both R and C as continuouslyinjected into a single space (namely here C) for the complex interpolation method to make sense.Concerning the Haagerup tensor product, for any sets I and J , we have a completely isometricidentity

OH(I) ⊗h OH(J) ≃ OH(I × J).

Finally, we should mention that OH is “homogeneous” (an o.s. E is called homogeneous if anylinear map u : E → E satisfies ‖u‖ = ‖u‖cb). While OH is unique, the class of homogeneous

48

Hilbertian operator spaces (which also includes R and C) is very rich and provides a very fruitfulsource of examples. See [120] for more on all this.

Remark. A classical application of Gaussian variables (resp. of the Khintchine inequality) in Lp isthe existence of an isometric (resp. isomorphic) embedding of ℓ2 into Lp for 0 < p 6= 2 < ∞ (ofcourse p = 2 is trivial). Thus it was natural to ask whether an analogous embedding (completelyisomorphic this time) holds for the operator space OH. The case 2 < p < ∞ was ruled out earlyon by Junge who observed that this would contradict (9.14). The crucial case p = 1 was solved byJunge in [60] (see [124] for a simpler proof). The case 1 < p ≤ 2 is included in Xu’s paper [156], aspart of more general embeddings. We refer the reader to Junge and Parcet’s [65] for a discussionof completely isomorphic embeddings of Lq into Lp (non-commutative) for 1 ≤ q < p < 2.

16. GT and random matrices

The notion of “exact operator space” originates in Kirchberg’s work on exactness for C∗-algebras.To abbreviate, it is convenient to introduce the exactness constant ex(E) of an operator space Eand to define “exact” operator spaces as those E such that ex(E) < ∞. We define first whendim(E) < ∞(16.1) dSK(E) = infdcb(E, F ) | n ≥ 1, F ⊂ Mn.By an easy perturbation argument, we have

(13.1)′ dSK(E) = infdcb(E, F ) | F ⊂ K(ℓ2),and this explains our notation. We then set

(16.2) ex(E) = supdSK(E1) | E1 ⊂ E dim(E1) < ∞and we call “exact” the operator spaces E such that ex(E) < ∞.

Note that if dim(E) < ∞ we have obviously ex(E) = dSK(E). Intuitively, the meaning of theexactness of E is a form of embeddability of E into the algebra K = K(ℓ2) of all compact operatorson ℓ2. But instead of a “global” embedding of E, it is a “local” form of embeddability that isrelevant here: we only ask for a uniform embeddability of all the finite dimensional subspaces.This local embeddability is of course weaker than the global one. By Theorem 12.3 and a simpleperturbation argument, any nuclear C∗-algebra A is exact and ex(A) = 1. A fortiori if A is nuclearand E ⊂ A then E is exact. Actually, if E is itself a C∗-algebra and ex(E) < ∞ then necessarilyex(E) = 1. Note however that a C∗-subalgebra of a nuclear C∗-algebra need not be nuclear itself.We refer the reader to [18] for an extensive treatment of exact C∗-algebras.A typical non-exact operator space (resp. C∗-algebra) is the space ℓ1 with its maximal o.s.s.described in Remark 13.13, (resp. the full C∗-algebra of any non-Abelian free group). Moreprecisely, it is known (see [123, p. 336]) that ex(ℓn

1 ) = dSK(ℓn1 ) ≥ n/(2

√n − 1).

In the Banach space setting, there are conditions on Banach spaces E ⊂ C(S1), F ⊂ C(S2)such that any bounded bilinear form ϕ : E×F → R or C satisfies the same factorization as in GT.But the conditions are rather strict (see Remark 6.2). Therefore the next result came as a surprise,because it seemed to have no Banach space analogue.

Theorem 16.1 ([66]). Let E, F be exact operator spaces. Let u : E → F ∗ be a c.b. map. LetC = ex(E) ex(F )‖u‖cb. Then for any n and any finite sets (x1, . . . , xn) in E, (y1, . . . , yn) in F wehave

∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ C(‖(xj)‖R + ‖(xj)‖C)(‖(yj)‖R + ‖(yj)‖C),(16.3)

49

and hence a fortiori

∑|〈uxj , yj〉| ≤ 2C(‖(xj)‖2

R + ‖(xj)‖2C)1/2(‖(yj)‖2

R + ‖(yj)‖2C)1/2.

Assume E ⊂ A, F ⊂ B (A, B C∗-algebra). Then there are states f1, f2 on A, g1, g2 on B such that

(16.4) ∀(x, y) ∈ E × F |〈ux, y〉| ≤ 2C(f1(x∗x) + f2(xx∗))1/2(g1(yy∗) + g2(y

∗y))1/2.

Moreover there is a linear map u : A → B∗ with ‖u‖ ≤ 4C satisfying

(16.5) ∀(x, y) ∈ E × F 〈ux, y〉 = 〈ux, y〉.

Although the original proof of this requires a much “softer” ingredient, it is easier to describethe argument using the following more recent (and much deeper) result, involving random matrices.This brings us back to Gaussian random variables. We will denote by Y (N) an N × N randommatrix, the entries of which, denoted by Y (N)(i, j) are independent complex Gaussian variableswith mean zero and variance 1/N so that E|Y N (i, j)|2 = 1/N for all 1 ≤ i, j ≤ N . It is known (dueto Geman, see [52]) that

(16.6) limN→∞

‖Y (N)‖MN= 2 a.s.

Let (Y(N)1 , Y

(N)2 , . . .) be a sequence of independent copies of Y (N), so that the family Y N

k (i, j) |k ≥ 1, 1 ≤ i, j ≤ N is an independent family of N(0, N−1) complex Gaussian.

In [52], a considerable strengthening of (16.6) is proved: for any finite sequence (x1, . . . , xn) in

Mk with n, k fixed, the norm of∑n

j=1 Y(N)j ⊗xj converges almost surely to a limit that the authors

identify using free probability. The next result, easier to state, is formally weaker than their result(note that it implies in particular lim supN→∞ ‖Y (N)‖MN

≤ 2).

Theorem 16.2 ([52]). Let E be an exact operator space. Then for any n and any (x1, . . . , xn) inE we have almost surely

(16.7) lim supN→∞

∥∥∥∑

Y(N)j ⊗ xj

∥∥∥MN (E)

≤ ex(E)(‖(xj)‖R + ‖(xj)‖C) ≤ 2 ex(E)‖(xj)‖RC .

Remark. The closest to (16.7) that comes to mind in the Banach space setting is the “type 2”property. But the surprise is that C∗-algebras (e.g. commutative ones) can be exact (and hencesatisfy (16.7)), while they never satisfy “type 2” in the Banach space sense.

Proof of Theorem 16.1. Let C1 = ex(E) ex(F ), so that C = C1‖u‖cb. We identify MN (E) withMN ⊗ E. By (13.11) applied with α = β = N−1/2I, we have a.s.

limN→∞

∣∣∣∣∣∣N−1

∑

ij

tr(Y(N)i Y

(N)∗j )〈uxi, yj〉

∣∣∣∣∣∣≤ ‖u‖cb lim

N→∞

∥∥∥∑

Y(N)i ⊗ xi

∥∥∥ limN→∞

∥∥∥∑

Y(N)j ⊗ yj

∥∥∥ ,

and hence by Theorem 16.2, we have

(16.8) limN→∞

∣∣∣∣∣∣

∑

ij

N−1 tr(Y(N)i Y

(N)∗j )〈uxi, yj〉

∣∣∣∣∣∣≤ C1‖u‖cb(‖(xi)‖R + ‖(xi)‖C)(‖(yj)‖R + ‖(yj)‖C).

50

But since

N−1 tr(Y(N)i Y

(N)∗j ) = N−1

N∑

r,s=1

Y(N)i (r, s)Y

(N)j (r, s)

and we may rewrite if we wish

Y(N)j (r, s) = N− 1

2 g(N)j (r, s)

where g(N)j (r, s) | j ≥ 1, 1 ≤ r, s ≤ N, N ≥ 1 is an independent family of complex Gaussian

variables each distributed like gC, we have by the strong law of large numbers

limN→∞

N−1 tr(Y(N)i Y

(N)∗j ) = lim

N→∞N−2

N∑

r,s=1

g(N)i (r, s)g

(N)j (r, s)

= E(gCi gC

j )

=

1 if i = j

0 if i 6= j.

Therefore, (16.8) yields∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ C1‖u‖cb(‖(xj)‖R + ‖(yj)‖C)(‖(yj)‖R + ‖(yj)‖C)

Remark 16.3. In the preceding proof we only used (13.11) with α, β equal to a multiple of theidentity, i.e. we only used the inequality

(16.9)

∣∣∣∣∣∣

∑

ij

1

Ntr(aibj)〈uxi, yj〉

∣∣∣∣∣∣≤ c

∥∥∥∑

ai ⊗ xi

∥∥∥MN (E)

∥∥∥∑

bj ⊗ yj

∥∥∥MN (F )

valid with c = ‖u‖cb when N ≥ 1 and∑

ai ⊗xi ∈ MN (E),∑

bj ⊗ yj ∈ MN (F ) are arbitrary. Notethat this can be rewritten as saying that for any x = [xij ] ∈ MN (E) and any y = [yij ] ∈ MN (F )we have

(16.10)

∣∣∣∣∣∣

∑

ij

1

N〈uxij , yji〉

∣∣∣∣∣∣≤ c‖x‖MN (E)‖y‖MN (F )

It turns out that (16.10) or equivalently (16.9) is actually a weaker property studied in [10] and[58] under the name of “tracial boundedness.” More precisely, in [10] u is called tracially bounded(in short t.b.) if there is a constant c such that (16.9) holds and the smallest such c is denoted by‖u‖tb. The preceding proof shows that (16.3) holds with ‖u‖tb in place of ‖u‖cb. This formulationof Theorem 16.1 is optimal. Indeed, the converse also holds:

Proposition 16.4. Let u : E → F ∗ be a linear map. Let C(u) be the best possible constant C suchthat (16.3) holds for all finite sequences (xj), (yj). Let C(u) = inf‖u‖ where the infimum runsover all u satisfying (16.5). Then

4−1C(u) ≤ C(u) ≤ C(u).

Moreover4−1‖u‖tb ≤ C(u) ≤ ex(E)ex(F )‖u‖tb.

In particular, for a linear map u : A → B∗ tracial boundedness is equivalent to ordinary bounded-ness, and in that case

4−1‖u‖ ≤ 4−1‖u‖tb ≤ C(u) ≤ ‖u‖.

51

Proof. That C(u) ≤ C(u) is a consequence of (7.1). That 4−1C(u) ≤ C(u) follows from a simplevariant of the argument for (iii) in Proposition 4.2 (the factor 4 comes from the use of (16.4)). Thepreceding remarks show that C(u) ≤ ex(E)ex(F )‖u‖tb.It remains to show 4−1‖u‖tb ≤ C(u). By (16.3) we have

∣∣∣∣∣∣

∑

ij

〈uxij , yji〉

∣∣∣∣∣∣≤ C(u)(‖(xij)‖R + ‖(xij)‖C)(‖(yji)‖R + ‖(yji)‖C).

Note that for any fixed i the sum Li =∑

j eij ⊗ xij satisfies ‖Li‖ = ‖∑j xijx

∗ij‖1/2 ≤ ‖x‖MN (E),

and hence ‖∑ij xijx

∗ij‖1/2 ≤ N1/2‖x‖MN (E). Similarly ‖∑

ij x∗ijxij‖1/2 ≤ N1/2‖x‖MN (E). This

implies(‖(xij)‖R + ‖(xij)‖C)(‖(yji)‖R + ‖(yji)‖C) ≤ 4N‖x‖MN (E)‖y‖MN (F )

and hence we obtain (16.10) with c ≤ 4C(u), which means that 4−1‖u‖tb ≤ C(u).

Remark 16.5. Consider Hilbert spaces H and K. We denote by Hr, Kr (resp. Hc, Kc) the associatedrow (resp. column) operator spaces. Then for any linear map u : Hr → K∗

r , u : Hr → K∗c ,

u : Hc → K∗r or u : Hc → K∗

c we have

‖u‖tb = ‖u‖.

Indeed this can be checked by a simple modification of the preceding Cauchy-Schwarz argument.

The next two statements follow from results known to Steen Thorbjørnsen since at least 1999(private communication). We present our own self-contained derivation of this, for use only in §20below.

Theorem 16.6. Consider independent copies Y ′i = Y

(N)i (ω′) and Y ′′

j = Y(N)j (ω′′) for (ω′, ω′) ∈

Ω × Ω. Then, for any n2-tuple of scalars (αij), we have

(16.11) limN→∞

∥∥∥∑

αijY(N)i (ω′) ⊗ Y

(N)j (ω′′)

∥∥∥M

N2

≤ 4(∑

|αij |2)1/2

for a.e. (ω′, ω′′) in Ω × Ω.

Proof. By (well known) concentration of measure arguments, it is known that (16.6) is essentiallythe same as the assertion that limN→∞ E‖Y (N)‖MN

= 2. Let ε(N) be defined by

E‖Y (N)‖MN= 2 + ε(N)

so that we know ε(N) → 0.Again by concentration of measure arguments (see e.g. [87, p. 41] or [113, (1.4) or chapter 2])

there is a constant β such that for any N ≥ 1 and and p ≥ 2 we have

(16.12) (E‖Y (N)‖pMN

)1/p ≤ E‖Y (N)‖MN+ β(p/N)1/2 ≤ 2 + ε(N) + β(p/N)1/2.

For any α ∈ Mn, we denote

Z(N)(α)(ω′, ω′′) =∑n

i,j=1αijY

(N)i (ω′) ⊗ Y

(N)j (ω′′).

Assume∑

ij |αij |2 = 1. We will show that almost surely

limN→∞ ‖Z(N)(α)‖ ≤ 4.

52

Note that by the invariance of (complex) canonical Gaussian measures under unitary transforma-tions, Z(N)(α) has the same distribution as Z(N)(uαv) for any pair u, v of n × n unitary matrices.Therefore, if λ1, . . . , λn are the eigenvalues of |α| = (α∗α)1/2, we have

Z(N)(α)(ω′, ω′′) dist=

∑n

j=1λjY

(N)j (ω′) ⊗ Y

(N)j (ω′′).

We claim that by a rather simple calculation of moments, one can show that for any even integerp ≥ 2 we have

(16.13) E tr|Z(N)(α)|p ≤ (E tr|Y (N)|p)2.

Accepting this claim for the moment, we find, a fortiori, using (16.12):

E‖Z(N)(α)‖pMN

≤ N2(E‖Y (N)‖pMN

)2 ≤ N2(2 + ε(N) + β(p/N)1/2)2p.

Therefore for any δ > 0

P‖Z(N)(α)‖MN> (1 + δ)4 ≤ (1 + δ)−pN2(1 + ε(N)/2 + (β/2)(p/N)1/2)2p.

Then choosing (say) p = 5(1/δ) log(N) we find

P‖Z(N)(α)‖MN> (1 + δ)4 ∈ O(N−2)

and hence (Borel–Cantelli) limN→∞‖Z(N)(α)‖MN≤ 4 a.s..

It remains to verify the claim. Let Z = ZN (α), Y = Y (N) and p = 2m. We have

E tr|Z|p = E tr(Z∗Z)m =∑

λi1λj1 . . . λimλjm(E tr(Y ∗i1Yj1 . . . Y ∗

imYjm))2.

Note that the only nonvanishing terms in this sum correspond to certain pairings that guaranteethat both λi1λj1 . . . λimλjm ≥ 0 and E tr(Y ∗

i1Yj1 . . . Y ∗

imYjm) ≥ 0. Moreover, by Holder’s inequality

for the trace we have

|E tr(Y ∗i1Yj1 . . . Y ∗

imYjm)| ≤ Π(E tr|Yik |p)1/pΠ(E tr|Yjk|p)1/p = E tr(|Y |p).

From these observations, we find

(16.14) E tr|Z|p ≤ E tr(|Y |p)∑

λi1λj1 . . . λimλjmE tr(Y ∗i1Yj1 . . . Y ∗

imYjm)

but the last sum is equal to E tr(|∑ λjYj |p) and since∑

λjYjdist= Y (recall

∑ |λj |2 =∑ |αij |2 = 1)

we haveE tr

(∣∣∣∑

αjYj

∣∣∣p)

= E tr(|Y |p),

and hence (16.14) implies (16.13).

Corollary 16.7. For any integer n and ε > 0 there is N and n-tuples of N × N matrices Y ′i |

1 ≤ i ≤ n and Y ′′j | 1 ≤ j ≤ n in MN such that

sup

∥∥∥∥∥∥

n∑

i,j=1

αijY′i ⊗ Y ′′

j

∥∥∥∥∥∥M

N2

∣∣∣ αij ∈ C,∑

ij|αij |2 ≤ 1

≤ (4 + ε)(16.15)

min

1

nN

∑n

1tr|Y ′

i |2,1

nN

∑n

1tr|Y ′′

j |2

≥ 1 − ε.(16.16)

53

Proof. Fix ε > 0. Let Nε be a finite ε-net in the unit ball of ℓn2

2 . By Theorem 16.16 we have foralmost (ω′, ω′′)

(16.17) limN→∞ supα∈Nε

∥∥∥∑n

i,j=1αijY

′i ⊗ Y ′′

j

∥∥∥M

N2

≤ 4,

We may pass from an ε-net to the whole unit ball in (16.17) at the cost of an extra factor (1+ε) andwe obtain (16.15). As for (16.16), the strong law of large numbers shows that the left side of (16.16)tends a.s. to 1. Therefore, we may clearly find (ω′, ω′′) satisfying both (16.15) and (16.16).

Remark 16.8. A close examination of the proof and concentration of measure arguments show thatthe preceding corollary holds with N of the order of c(ε)n2.

Remark 16.9. Using the well known “contraction principle” that says that the variables (εj) aredominated by either (gR

j ) or (gCj ), it is easy to deduce that Corollary 16.7 is valid for matrices

Y ′i , Y ′′

j with entries all equal to ±N−1/2, with possibly a different numerical constant in place of4. Analogously, using the polar factorizations Y ′

i = U ′i |Y ′

i |, Y ′′j = U”j |Y ′′

j | and noting that allthe factors U ′

i , |Y ′i |, U”j , |Y ′′

j | are independent, we can also (roughly by integrating over the moduli|Y ′

i |, |Y ′′j |) obtain Corollary 16.7 with unitary matrices Y ′

i , Y ′′j , with a different numerical constant

in place of 4.

17. GT for exact operator spaces

We now abandon tracially bounded maps and turn to c.b. maps. In [126], the following fact playsa crucial role.

Lemma 17.1. Assume E and F are both exact. Let A, B be C∗-algebras, with either A or BQWEP. Then for any u in CB(E, F ∗) the bilinear map ΦA,B introduced in Proposition 13.9 satisfies

‖ΦA,B : A ⊗min E × B ⊗min F → A ⊗max B‖ ≤ ex(E)ex(F )‖u‖cb.

Proof. Assume A = W/I with W WEP. We also have, assuming say B separable that B = C/Jwith C LLP (e.g. C = C∗(F∞)). The exactness of E, F gives us

A ⊗min E = (W ⊗min E)/(I ⊗min E) and B ⊗min F = (C ⊗min F )/(J ⊗min F ).

Thus we are reduced to show the Lemma with A = W and B = C. But then by Kirchberg’sTheorem 12.7, we have A ⊗min B = A ⊗max B (with equal norms).

We now come to the operator space version of GT of [126]. We seem here to repeat the settingof Theorem 16.1. Note however that, by Remark 16.3, Theorem 16.1 gives a characterization oftracially bounded maps u : E → F ∗, while the next statement characterizes completely boundedones.

Theorem 17.2 ([126]). Let E, F be exact operator spaces. Let u ∈ CB(E, F ∗). Assume ‖u‖cb ≤ 1.Let C = ex(E)ex(F ). Then for any finite sets (xj) in E, (yj) in F we have

(17.1)∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ 2C(‖(xj)‖R‖(yj)‖C + ‖(xj)‖C‖(yj)‖R).

Equivalently, assuming E ⊂ A, F ⊂ B there are states f1, f2 on A g1, g2 on B such that

(17.2) ∀(x, y) ∈ E × F |〈ux, y〉| ≤ 2C((f1(xx∗)g1(y∗y))1/2 + (f2(x

∗x)g2(yy∗))1/2).

Conversely if this holds for some C then ‖u‖cb ≤ 4C.

54

Proof. The main point is (17.1). The equivalence of (17.1) and (17.2) is explained below in Propo-sition 18.2. To prove (17.1), not surprisingly, Gaussian variables reappear as a crucial ingredient.But it is their analogue in Voiculescu’s free probability theory (see [151]) that we need, and actuallywe use them in Shlyakhtenko’s generalized form. To go straight to the point, what this theory doesfor us is this: Let (tj) be an arbitrary finite set with tj > 0. Then we can find operators (aj) and(bj) on a (separable) Hilbert space H and a unit vector ξ such that

(i) aibj = bjai and a∗i bj = bja∗i for all i, j.

(ii) 〈aibjξ, ξ〉 = δij .

(iii) For any (xj) and (yj) in B(K) (K arbitrary Hilbert)

∥∥∥∑

aj ⊗ xj

∥∥∥min

≤ ‖(tjxj)‖R + ‖(t−1j xj)‖C

∥∥∥∑

bj ⊗ yj

∥∥∥min

≤ ‖(tjyj)‖R + ‖(t−1j yj)‖C .

(iv) The C∗-algebra generated by (aj) is QWEP.

With this crucial ingredient, the proof of (17.1) is easy to complete: By Lemma 17.1, (ii) implies

∣∣∣∑

〈uxj , yj〉∣∣∣ =

∣∣∣∣∣∣

∑

i,j

〈uxi, yj〉〈aibjξ, ξ〉

∣∣∣∣∣∣

≤∥∥∥∑

〈uxi, yj〉ai ⊗ bj

∥∥∥max

≤ C∥∥∥∑

aj ⊗ xj

∥∥∥min

∥∥∥∑

bj ⊗ yj

∥∥∥min

and hence (iii) yields

∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ C(‖(tjxj)‖R + ‖(t−1

j xj)‖C)(‖(tjyj)‖R + ‖(t−1j yj)‖C).

A fortiori (here we use the elementary inequality (a + b)(c + d) ≤ 2(a2 + b2)1/2(c2 + d2)1/2 ≤s2(a2 + b2) + s−2(c2 + d2) valid for non-negative numbers a, b, c, d and s > 0)

∣∣∣∑

〈uxj , yj〉∣∣∣ ≤ C

(s2

∥∥∥∑

t2jx∗jxj

∥∥∥ + s2∥∥∥∑

t−2j xjx

∗j

∥∥∥ + s−2∥∥∥∑

t2jy∗j yj

∥∥∥ + s−2∥∥∥∑

t−2j yjy

∗j

∥∥∥)

.

By the Hahn–Banach argument (see §4) this implies the existence of states f1, f2, g1, g2 such that

∀(x, y) ∈ E × F |〈ux, y〉| ≤ C(s2t2f1(x∗x) + s2t−2f2(xx∗) + s−2t2g2(y

∗y) + s−2t−2g1(yy∗)).

Then taking the infimum over all s, t > 0 we obtain (17.2) and hence (17.1).

We now describe briefly the generalized complex free Gaussian variables that we use, followingVoiculescu’s and Shlyakhtenko’s ideas. One nice realization of these variables is on the Fock space.But while it is well known that Gaussian variables can be realized using the symmetric Fock space,here we need the “full” Fock space, as follows: We assume H = ℓ2(I), and we define

F (H) = C ⊕ H ⊕ H⊗2 ⊕ · · ·

55

where H⊗n denotes the Hilbert space tensor product of n-copies of H. As usual one denotes by Ωthe unit in C viewed as sitting in F (H) (“vacuum vector”).

Given h in H, we denote by ℓ(h) (resp. r(h)) the left (resp. right) creation operator, definedby ℓ(h)x = h ⊗ x (resp. r(h)x = x ⊗ h). Let (ej) be the canonical basis of ℓ2(I). Then the familyℓ(ej) + ℓ(ej)

∗ | j ∈ I is a “free semi-circular” family ([151]). This is the free analogue of (gRj ), so

we refer to it instead as a “real free Gaussian” family. But actually, we need the complex variant.We assume I = J × 1, 2 and then for any j in J we set

Cj = ℓ(e(j,1)) + ℓ(e(j,2))∗.

The family (Cj) is a “free circular” family (cf. [151]), but we refer to it as a “complex free Gaussian”family. The generalization provided by Shlyakhtenko is crucial for the above proof. This uses thepositive weights (tj) that we assume fixed. One may then define

aj = tjℓ(e(j,1)) + t−1j ℓ(e(j,2))

∗(17.3)

bj = tjr(e(j,2)) + t−1j r(e(j,1))

∗,(17.4)

and set ξ = Ω. Then it is not hard to check that (i), (ii) and (iii) hold. In sharp contrast, (iv) ismuch more delicate but it follows from known results from free probability, namely the existenceof asymptotic “matrix models” for free Gaussian variables or their generalized versions (17.3) and(17.4).

Corollary 17.3. An operator space E is exact as well as its dual E∗ iff there are Hilbert spacesH, K such that E ≃ Hr ⊕ Kc completely isomorphically.

Proof. By Theorem 17.2 (and Proposition 18.2 below) the identity of E factors through Hr ⊕Kc.By a remarkable result due to Oikhberg [101] (see [124] for a simpler proof), this can happen only ifE itself is already completely isomorphic to Hr ⊕Kc for subspaces H⊂ H and K ⊂ K. This provesthe only if part. The “if part” is obvious since Hr, Kc are exact and (Hr)

∗ ≃ Hc, (Kc)∗ ≃ Kr.

We postpone to the next chapter the discussion of the case when E = A and F = B in Theorem17.2. We will also indicate there a new proof of Theorem 17.2 based on Theorem 14.1 and the morerecent results of [49].

18. GT for operator spaces

In the Banach space case, GT tells us that bounded linear maps u : A → B∗ (A, B C∗-algebras,commutative or not, see §2 and §8) factor (boundedly) through a Hilbert space. In the operatorspace case, we consider c.b. maps u : A → B∗ and we look for a c.b. factorization through someHilbertian operator space. It turns out that, if A or B is separable, the relevant Hilbertian spaceis the direct sum

R ⊕ C

of the row and column spaces introduced in §13, or more generally the direct sum

Hr ⊕ Kc

where H, K are arbitrary Hilbert spaces.In the o.s. context, it is more natural to define the direct sum of two operator spaces E, F as the

“block diagonal” direct sum, i.e. if E ⊂ A and F ⊂ B we embed E ⊕ F ⊂ A⊕B and equip E ⊕ F

56

with the induced o.s.s. Note that for all (x, y) ∈ E ⊕ F we have then ‖(x, y)‖ = max‖x‖, ‖y‖.Therefore the spaces R⊕C or Hr ⊕Kc are not isometric but only

√2-isomorphic to Hilbert space,

but this will not matter much.In analogy with (16.3), for any linear map u : E → F between operator spaces we denote by

γr⊕c(u)) the constant of factorization of u through a space of the form Hr ⊕ Kc. More precisely,we set

γr⊕c(u) = inf‖u1‖cb‖u2‖cbwhere the infimum runs over all possible Hilbert spaces H, K and all factorizations

Eu1−→ Z

u2−→ F

of u through Z with Z = Hr ⊕ Kc. Let us state a first o.s. version of GT.

Theorem 18.1. Let A, B be C∗-algebra. Then any c.b. map u : A → B∗ factors through a spaceof the form Hr ⊕ Kc for some Hilbert space H, K. If A or B is separable, we can replace Hr ⊕ Kc

simply by R ⊕ C. More precisely, for any such u we have γr⊕c(u) ≤ 2‖u‖cb.

Curiously, the scenario of the non-commutative GT repeated itself: this was proved in [126]assuming that either A or B is exact, or assuming that u is suitably approximable by finite ranklinear maps. These restrictions were removed in the recent paper [49] that also obtained betterconstants. A posteriori, this means (again!) that the approximability assumption of [126] for c.b.maps from A to B∗ holds in general. The case of exact operator subspaces E ⊂ A, F ⊂ B (i.e.Theorem 17.2) a priori does not follow from the method in [49] but, in the second proof of Theorem17.2 given at the end of this section, we will show that it can also be derived from the ideas of [49]using Theorem 16.1 in place of Theorem 7.1.

Just like in the classical GT, the above factorization is equivalent to a specific inequality whichwe now describe, following [126].

Proposition 18.2. Let E ⊂ A, F ⊂ B be operator spaces and let u : E → F ∗ be a linear map(equivalently we may consider a bilinear form on E × F ). The following assertions are equivalent:

(i) For any finite sets (xj), (yj) in E, F respectively and for any number tj > 0 we have

∣∣∣∑

〈uxjyj〉∣∣∣ ≤ (‖(xj)‖R‖(yj)‖C + ‖(tjxj)‖C‖(t−1

j yj)‖R).

(ii) There are states f1, f2 on A, g1, g2 on B such that

∀(x, y) ∈ E × F |〈ux, y〉| ≤ (f1(xx∗)g1(y∗y))1/2 + (f2(x

∗x)g2(yy∗))1/2.

(iii) There is a decomposition u = u1 + u2 with maps u1 : E → F ∗ and u2 : E → F ∗ such that

∀(x, y) ∈ E × F |〈u1x, y〉| ≤ (f1(xx∗)g1(y∗y))1/2 and |〈u2x, y〉| ≤ (f2(x

∗x)g2(yy∗))1/2.

(iv) There is a decomposition u = u1 + u2 with maps u1 : E → F ∗ and u2 : E → F ∗ such thatγr(u1) ≤ 1 and γc(u2) ≤ 1.

In addition, the bilinear form associated to u on E×F extends to one on A×B that still satisfies (i).Moreover, these conditions imply γr⊕c(u) ≤ 2, and conversely γr⊕c(u) ≤ 1 implies these equivalentconditions.

57

Proof. (Sketch) The equivalence between (i) and (ii) is proved by the Hahn–Banach type argumentin §4. (ii) ⇒ (iii) (with the same states) requires a trick of independent interest, due to the author,see [155, Prop. 5.1]. Assume (iii). Then by Theorem 14.1 and Remark 14.2, we have γr(u1) ≤ 1and γc(u2) ≤ 1. By the triangle inequality this implies (ii). The extension property follows from(iii) in Theorem 4.2. The last assertion is easy by Theorem 14.1 and Remark 14.2.

The key ingredient for the proof of Theorem 18.1 is the “Powers factor” Mλ, i.e. the von Neu-mann algebra associated to the state

ϕλ =⊗

N

( λ1+λ 0

0 11+λ

)

on the infinite tensor product of 2 × 2 matrices. If λ 6= 1 the latter is of “ type III”, i.e. does notadmit any kind of trace. We will try to describe “from scratch” the main features that are neededfor our purposes in a somewhat self-contained way, using as little von Neumann Theory as possible.Here (and throughout this section) λ will be a fixed number such that

0 < λ < 1.

For simplicity of notation, we will drop the superscript λ and denote simply ϕ, M, N, . . . instead ofϕλ, Mλ, Nλ, . . . but the reader should recall these do depend on λ. The construction of Mλ (basedon the classical GNS construction) can be outlined as follows:

With M2 denoting the 2 × 2 complex matrices, let An = M⊗n2 equipped with the C∗-norm

inherited from the identification with M2n . Let A = ∪An where we embed An into An+1 via theisometric map x → x⊗ 1. Clearly, we may equip A with the norm inherited from the norms of thealgebras An.

Let ϕn = ψ ⊗ · · · ⊗ ψ (n-times) with ψ =

( λ1+λ 0

0 11+λ

). We define ϕ ∈ A∗ by

∀a ∈ A ϕ(a) = limn→∞

tr(ϕna)

where the limit is actually stationary, i.e. we have ϕ(a) = tr(ϕna) ∀a ∈ An (here the trace is meantin M⊗n

2 ≃ M2n). We equip A with the inner product:

∀a, b ∈ A 〈a, b〉 = ϕ(b∗a).

The space L2(ϕ) is then defined as the Hilbert space obtained from (A, 〈·, ·〉) after completion.We observe that A acts on L2(ϕ) by left multiplication, since we have ‖ab‖L2(ϕ) ≤ ‖a‖A‖b‖L2(ϕ)

for all a, b ∈ A. So from now on we view

A ⊂ B(L2(ϕ)).

We then let M be the von Neumann algebra generated by A, i.e. we set M = A′′ (bicommutant).Recall that, by classical results, the unit ball of A is dense in that of M for either the weak orstrong operator topology (wot and sot in short).

Let L denote the inclusion map into B(L2(ϕ)). Thus

(18.1) L : M → B(L2(ϕ))

is an isometric ∗-homomorphism extending the action of left multiplication. Indeed, let b → bdenote the dense range inclusion of A into L2(ϕ). Then we have

∀a ∈ A ∀b ∈ A L(a)b =

.︷︸︸︷ab .

58

Let ξ = 1. Note L(a)ξ = a and also 〈L(a)ξ, ξ〉 = ϕ(a) for all a in A. Thus we can extend ϕ to thewhole of M by setting

(18.2) ∀a ∈ M ϕ(a) = 〈L(a)ξ, ξ〉.

We wish to also have an action of M analogous to right multiplication. Unfortunately, when λ 6= 1,ϕ is not tracial and hence right multiplication by elements of M is unbounded on L2(ϕ). Thereforewe need a modified version of right multiplication, as follows.

For any a, b in A, let n be such that a, b ∈ An, we define

(18.3) R(a)b =

.︷ ︸︸ ︷b(ϕ1/2

n aϕ− 1

2n ).

Note that this does not depend on n (indeed ϕ1/2n+1(a ⊗ 1)ϕ

− 12

n+1 = (ϕ1/2n aϕ

− 12

n ) ⊗ 1).A simple verification shows that

∀a, b ∈ A ‖R(a)b‖L2(ϕ) ≤ ‖a‖A‖b‖L2(ϕ),

and hence this defines R(a) ∈ B(L2(ϕ)) by density of A in L2(ϕ). Note that

(18.4) ∀a ∈ A 〈R(a)ξ, ξ〉 = 〈L(a)ξ, ξ〉 = ϕ(a) and R(a∗) = R(a)∗.

Using this together with the wot-density of the unit ball of A in that of M , we can extend a 7→ R(a)to the whole of M . Moreover, by (18.4) we have (note R(a)R(a∗) = R(a∗a))

(18.5) ∀a ∈ M 〈R(a∗)ξ, R(a∗)ξ〉 = 〈L(a)ξ, L(a)ξ〉 = ϕ(a∗a).

Since left and right multiplication obviously commute we have

∀a1, a2 ∈ A R(a1)L(a2) = L(a2)R(a1),

and hence also for all a1, a2 ∈ M . Thus we obtain a ∗-anti-homomorphism

R : M → B(L2(ϕ)),

i.e. such that R(a1a2) = R(a2)R(a1) ∀a1, a2 ∈ M . Equivalently, let Mop denote the von Neumannalgebra that is “opposite” to M , i.e. the same normed space with the same involution but withreverse product (i.e. a · b = ba by definition). If M ⊂ B(ℓ2), Mop can be realized concretely as thealgebra tx | x ∈ M. Then R : Mop → B(L2(ϕ)) is a bonafide ∗-homomorphism. We may clearlydo the following identifications

M2n ≃ An ≃ L(An) ≃ M2 ⊗ · · · ⊗ M2 ⊗ 1 ⊗ · · ·

where M2 is repeated n times and 1 infinitely many times.In particular we view An ⊂ M , so that we also have L2(An, ϕ|An

) ⊂ L2(ϕ) and ϕ|An= ϕn. Let

Pn : L2(ϕ) → L2(An, ϕ|An) be the orthogonal projection.

Then for any a in A, say a = a1 ⊗ · · · ⊗ an ⊗ an+1 ⊗ · · · , we have

Pn(a) = a1 ⊗ · · · ⊗ an ⊗ 1 · · · (ψ(an+1)ψ(an+2) · · · ),

This behaves like a conditional expectation, e.g. for any a, b ∈ An, we have PnL(a)R(b) =L(a)R(b)Pn. Moreover, since the operator PnL(a)|L2(An,ϕ|An

) commutes with right multiplications,

59

for any a ∈ M , there is a unique element an ∈ An such that ∀x ∈ L2(An, ϕ|An), PnL(a)x = L(an)x.

We will denote En(a) = an. Note that En(a∗) = En(a)∗. For any x ∈ L2(ϕ), by a den-sity argument, we obviously have Pn(x) → x in L2(ϕ). Note that for any a ∈ M , we haveL(an)ξ = Pn(L(a)ξ) → L(a)ξ in L2(ϕ), thus, using (18.5) we find

(18.6) ‖L(a − an)ξ‖L2(ϕ) = ‖R(a∗ − a∗n)ξ‖L2(ϕ) → 0.

Note that, if we identify (completely isometrically) the elements of An with left multiplications onL2(An, ϕ|An

), we may write En(.) = PnL(.)|L2(An,ϕ|An). The latter shows (by Theorem 13.3) that

En : M → An is a completely contractive mapping. Thus we have

(18.7) ∀n ≥ 1 ‖En : M → An‖cb ≤ 1.

Obviously we may write as well:

(18.8) ∀n ≥ 1 ‖En : Mop −→ Aopn ‖cb ≤ 1.

(The reader should observe that En and Pn are essentially the same map, since we have

.︷ ︸︸ ︷En(a) =

Pn(a) , or equivalently L(En(a))ξ = Pn(L(a)ξ) but the multiple identifications may be confusing.)Let N ⊂ M denote the ϕ-invariant subalgebra, i.e.

N = a ∈ M | ϕ(ax) = ϕ(xa) ∀x ∈ M.

Obviously, N is a von Neumann subalgebra of M . Moreover ϕ|N is a faithful tracial state (andactually a “vector state” by (18.2)), so N is a finite von Neumann algebra.We can now state the key analytic ingredient for the proof of Theorem 18.1.

Lemma 18.3. Let Φ: M × Mop → C and Φn : M × Mop → C be the bilinear forms defined by

Φ(a, b) = 〈L(a)R(b)ξ, ξ〉∀a, b ∈ M

Φn(a, b) = tr(En(a)ϕ1/2n En(b)ϕ1/2

n )

where tr is the usual trace on M⊗n2 ≃ An ⊂ M .

(i) For any a, b in M , we have

(18.9) ∀a, b ∈ M limn→∞

Φn(a, b) = Φ(a, b).

(ii) For any u in U(N) (the set of unitaries in N) we have

(18.10) ∀a, b ∈ M Φ(uau∗, ubu∗) = Φ(a, b)

(iii) For any u in M , let α(u) : M → M denote the mapping defined by α(u)(a) = uau∗. Let C =α(u) | u ∈ U(n). There is a net of mappings αi in conv(C) such that ‖αi(v)−ϕλ(v)1‖ → 0for any v in N .

(iv) For any q ∈ Z, there exists cq in M (actually in A|q|) such that c∗qcq and cqc∗q both belong to

N and such that:

(18.11) ϕ(c∗qcq) = λ−q/2, ϕ(cqc∗q) = λq/2 and Φ(cq, c

∗q) = 1.

60

We give below a direct proof of this Lemma that is “almost” self-contained (except for the useDixmier’s classical averaging theorem and a certain complex interpolation argument). But we firstshow how this Lemma yields Theorem 18.1.

Proof of Theorem 18.1. Consider the bilinear form

Φ: M ⊗min A × Mop ⊗min B −→ C

defined (on the algebraic tensor product) by

Φ(a ⊗ x, b ⊗ y) = Φ(a, b)〈ux, y〉.

We define Φn similarly. We claim that Φ is bounded with

‖Φ‖ ≤ ‖u‖cb.

Indeed, by Proposition 13.9 (note the transposition sign there), and by (18.7) and (18.8) we have

∣∣∣Φn

(∑ar ⊗ xr,

∑bs ⊗ ys

)∣∣∣ ≤ ‖u‖cb

∥∥∥∑

En(ar) ⊗ xr

∥∥∥M2n(A)

∥∥∥∑

tEn(bs) ⊗ ys

∥∥∥M2n(B)

≤ ‖u‖cb

∥∥∥∑

ar ⊗ xr

∥∥∥M⊗minA

∥∥∥∑

bs ⊗ ys

∥∥∥Mop⊗minB

,

and then by (18.9) we obtain the announced claim ‖Φ‖ ≤ ‖u‖cb.But now by Theorem 6.1, there are states f1, f2 on M ⊗min A, g1, g2 on Mop ⊗min B such that

∀X ∈ M ⊗ A,∀Y ∈ M ⊗ B |Φ(X, Y )| ≤ ‖u‖cb(f1(X∗X) + f2(XX∗))1/2(g2(Y

∗Y ) + g1(Y Y ∗))1/2

In particular, we may apply this when X = cq ⊗ x and Y = c∗q ⊗ y. Recall that by (18.11)Φ(cq, c

∗q) = 1. We find

|〈ux, y〉||Φ(cq, c∗q)| ≤ ‖u‖cb(f1(c

∗qcq ⊗ x∗x) + f2(cqc

∗q ⊗ xx∗))1/2(g2(c

∗qcq ⊗ y∗y) + g1(cqc

∗q ⊗ yy∗))1/2.

But then by (18.10) we have for any i

Φ(αi(cq), αi(c∗q)) = Φ(cq, c

∗q)

and since c∗qcq, cqc∗q ∈ N , we know by Lemma 18.3 and (18.11) that αi(c

∗qcq) → ϕ(c∗qcq)1 = λ−q/21

while αi(cqc∗q) → ϕ(cqc

∗q)1 = λq/21. Clearly this implies αi(c

∗qcq) ⊗ x∗x → λ−q/21 ⊗ x∗x and

αi(cqc∗q) ⊗ xx∗ → λq/21 ⊗ xx∗ and similarly for y. It follows that if we denote

∀x ∈ A fk(x) = fk(1 ⊗ x) and ∀y ∈ B gk(y) = gk(1 ⊗ y)

then fk, gk are states on A, B respectively such that(18.12)∀(x, y) ∈ A × B |〈ux, y〉| ≤ ‖u‖cb(λ

−q/2f1(x∗x) + λq/2f2(xx∗))1/2(λ−q/2g2(y

∗y) + λq/2g1(yy∗))1/2.

Then we find|〈ux, y〉|2 ≤ ‖u‖2

cb(f1(x∗x)g1(yy∗) + f2(xx∗)g2(y

∗y) + δq(λ))

where δq(λ) = λ−qβ + λqα where we set

β = f1(x∗x)g2(y

∗y) and α = f2(xx∗)g1(yy∗).

61

But an elementary calculation shows that (here we crucially uses that λ < 1)

infq∈Z

δq(λ) ≤ (λ1/2 + λ−1/2)√

αβ

so after minimizing over q ∈ Z we find

|〈ux, y〉| ≤ ‖u‖cb(f1(x∗x)g1(yy∗) + f2(xx∗)g2(y

∗y) + (λ1/2 + λ−1/2)√

αβ)1/2

and since C(λ) = (λ1/2 + λ−1/2)/2 ≥ 1 we obtain ∀x ∈ A ∀y ∈ B

|〈ux, y〉| ≤ ‖u‖cbC(λ)1/2((f1(x

∗x)g1(yy∗))1/2 + (f2(xx∗)g2(y∗y))1/2

).

To finish, we note that C(λ) → 1 when λ → 1, and, by pointwise compactness, we may assumethat the above states f1, f2, g1, g2 that (implicitly) depend on λ are also converging pointwise whenλ → 1. Then we obtain the announced result, i.e. the last inequality holds with the constant 1 inplace of C(λ).

Proof of Lemma 18.3. (i) Let an = En(a), bn = En(b), (a, b ∈ M). We know that an → a, bn → b

and also

.︷︸︸︷a∗n →

.︷︸︸︷a∗ . Therefore we have by (18.6)

Φn(a, b) = 〈L(an)ξ, R(b∗n)ξ〉 → 〈L(a)ξ, R(b∗)ξ〉 = Φ(a, b).

(iii) Since ϕ|N is a (faithful normal) tracial state on N , N is a finite factor, so this follows fromDixmier’s classical approximation theorem ([26] or [68, p. 520]).(iv) The case q = 0 is trivial, we set c0 = 1. Let c = (1 + λ)1/2λ−1/4e12. We then set, for q ≥ 1,cq = c⊗· · ·⊗c⊗1 · · · where c is repeated q times and for q < 0 we set cq = (c−q)

∗. The verificationof (18.11) then boils down to the observation that

ψ(e∗12e12) = (1 + λ)−1, ψ(e12e∗12) = λ(1 + λ)−1 and tr(ψ1/2e12ψ

1/2e21) = (1 + λ)−1λ1/2.

(ii) By polarization, it suffices to show(18.10) for b = a∗. The proof can be completed usingproperties of self-polar forms (one can also use the Pusz-Woronowicz “purification of states”ideas). We outline an argument based on complex interpolation. Consider the pair of Banachspaces (L2(ϕ), L2(ϕ)†) where L2(ϕ)† denotes the completion of A with respect to the norm x 7→(ϕ(xx∗))1/2. Clearly we have natural continuous inclusions of both these spaces into M∗ (givenrespectively by x 7→ xϕ and x 7→ ϕx). Let us denote

Λ(ϕ) = (L2(ϕ), L2(ϕ)†)1/2.

Similarly we denote for any n ≥ 1

Λ(ϕn) = (L2(An, ϕn), L2(An, ϕn)†)1/2.

By a rather easy elementary (complex variable) argument one checks that for any an ∈ An we have

‖an‖2Λ(ϕn) = tr(ϕ1/2

n anϕ1/2n a∗n) = Φn(an, a∗n).

By (18.5) and (18.6), we know that for any a ∈ M , if an = En(a), then a − an tends to 0 inboth spaces (L2(ϕ), L2(ϕ)†), and hence in the interpolation space Λ(ϕ). In particular we have‖a‖2

Λ(ϕ) = limn ‖an‖2Λ(ϕ). Since En(a∗) = En(a)∗ for any a ∈ M , En defines a norm one projection

62

simultaneously on both spaces (L2(ϕ), L2(ϕ)†), and hence in Λ(ϕ). This implies that ‖an‖2Λ(ϕ) =

‖an‖2Λ(ϕn) = Φn(an, a∗n). Therefore, for any a ∈ M we find using (i)

(18.13) ‖a‖2Λ(ϕ) = Φ(a, a∗).

Now for any unitary u in N , for any a ∈ M , we obviously have ‖uau∗‖L2(ϕ) = ‖a‖L2(ϕ) and‖uau∗‖L2(ϕ)† = ‖a‖L2(ϕ)† . By the basic interpolation principle (applied to a 7→ uau∗ and its inverse)this implies ‖a‖Λ(ϕ) = ‖uau∗‖Λ(ϕ). Thus by (18.13) we conclude Φ(a, a∗) = Φ(uau∗, ua∗u∗), and(ii) follows.

Second proof of Theorem 17.2. (jointly with Mikael de la Salle). Consider a c.b. map u : E → F ∗.Consider the bilinear form

θn : An ⊗min E ×Aopn ⊗min F → C

defined by θn(a ⊗ x, b ⊗ y) = Φn(a, b)〈ux, y〉. Note that An ⊗min E and Aopn ⊗min F are exact with

constants at most respectively ex(E) and ex(F ). Therefore Theorem 16.1 tells us that θn satisfies

(16.4) with states f(n)1 , f

(n)2 , g

(n)1 , g

(n)2 respectively on M ⊗min A and Mop ⊗min B. Let Φn and Φ be

the bilinear forms defined like before but this time on M ⊗min E ×Mop ⊗min F . Arguing as beforeusing (18.7) and (18.8) we find that ∀X ∈ M ⊗ E ∀Y ∈ Mop ⊗ F we have

(18.14) |Φn(X, Y )| ≤ 2C(f(n)1 (X∗X) + f

(n)2 (XX∗))1/2(g

(n)2 (Y ∗Y ) + g

(n)1 (Y Y ∗))1/2.

Passing to a subsequence, we may assume that f(n)1 , f

(n)2 , g

(n)1 , g

(n)2 all converge pointwise to states

f1, f2, g1, g2 respectively on M ⊗min A and Mop ⊗min B. Passing to the limit in (18.14) we obtain

|Φ(X, Y )| ≤ 2C(f1(X∗X) + f2(XX∗))1/2(g2(Y

∗Y ) + g1(Y Y ∗))1/2.

We can then repeat word for word the end of the proof of Theorem 18.1 and we obtain (17.2).

Remark. In analogy with Theorem 18.1, it is natural to investigate whether the Maurey factorizationdescribed in § 10 has an analogue for c.b. maps from A to B∗ when A, B are non-commutativeLp-space and 2 ≤ p < ∞. This program was completed by Q. Xu in [155]. Note that thisrequires proving a version of Khintchine’s inequality for “generalized circular elements”, i.e. fornon-commutative “random variables” living in Lp over a type III von Neumann algebra. Roughlythis means that all the analysis has to be done “without any trace”!

19. GT and Quantum mechanics: EPR and Bell’s inequality

In 1935, Einstein, Podolsky and Rosen (EPR in short) published a famous article vigorously criti-cizing the foundations of quantum mechanics (QM in short). They pushed forward the alternativeidea that there are, in reality, “hidden variables” that we cannot measure because our technicalmeans are not yet powerful enough, but that the statistical aspects of quantum mechanics can bereplaced by this concept.

In 1964, J.S. Bell observed that the hidden variables theory could be put to the test. Heproposed an inequality (now called “Bell’s inequality”) that is a consequence of the hidden variablesassumption.

Clauser, Holt, Shimony and Holt (CHSH, 1969), modified the Bell inequality and suggested thatexperimental verification should be possible. Many experiments later, there seems to be consensusamong the experts that the Bell-CHSH inequality is violated, thus essentially the EPR “hiddenvariables” theory is invalid, and in fact the measures tend to agree with the predictions of QM.

63

We refer the reader to Alain Aspect’s papers [6, 7] for an account of the experimental saga, andfor relevant physical background to the books [110, Chapter 6] or [8, Chapter 10]. For simplicity,we will not discuss here the concept of “locality” or “local realism” related to the assumption thatthe observations are independent (see [6]). See [8, p. 196] for an account of the Bohm theory wherenon-local hidden variables are introduced in order to reconcile theory with experiments.

In 1980 Tsirelson observed that GT could be interpreted as giving an upper bound for theviolation of a (general) Bell inequality, and that the violation of Bell’s inequality is closely relatedto the assertion that KG > 1! He also found a variant of the CHSH inequality (now called“Tsirelson’s bound”), see [146, 147, 148, 150, 31].

The relevant experiment can be schematically described by the following diagram.

We have a source that produces two particles (photons) issued from the split of a single particlewith zero spin. The new particles are sent horizontally in opposite directions toward two observers,A (Alice) and B (Bob) that are assumed far apart so whatever measuring they do does not influencethe result of the other. If they both use a detector placed exactly in horizontal position, the arrivingparticle will be invariably read as precisely opposite spins +1 and −1 since originally the spin waszero. So with certainty the product result will be −1. However, assume now that Alice and Bobcan use their detector in different angular positions, that we designate by i (1 ≤ i ≤ n, so thereare n positions of their device). Let Ai = ±1 denote the result of Alice’s detector and Bi denotethe result of Bob’s. Again we have Bi = −Ai Assume now that Bob uses a detector in a positionj, while Alice uses position i 6= j. Then the product AiBj is no longer deterministic, the result israndomly equal to ±1, but its average, i.e. the covariance of Ai and Bj can be measured (indeed,A and B can repeat the same measurements many times and confront their results, not necessarilyinstantly, say by phone afterwards, to compute this average). We will denote it by ξij . Here is anoutline of Bell’s argument:According to the EPR theory, there are “hidden variables” that can be denoted by a single one λand a probability distribution ρ(λ)dλ so that the covariance of Ai and Bj is

ξij =

∫Ai(λ)Bj(λ)ρ(λ)dλ.

We will now fix a real matrix [aij ] (to be specified later). Then for any ρ we have

|∑

aijξij | ≤ HV (a)max = supφi=±1ψj=±1

|∑

aijφiψj | = ‖a‖∨.

Equivalently HV (a)max denotes the maximum value of |∑ aijξij | over all possible distributions ρ.But Quantum Mechanics predicts

ξij = tr(ρAiBj)

where Ai, Bj are self-adjoint unitary operators on H (dim(H) < ∞) with spectrum in ±1 suchthat AiBj = BjAi (this reflects the separation of the two observers) and ρ is a non-commutative

64

probability density, i.e. ρ ≥ 0 is a trace class operator with tr(ρ) = 1. This yields

|∑

aijξij | ≤ QM(a)max = supρ,Ai,Bj

|tr(ρ∑

aijAiBj)| = supx∈BH ,Ai,Bj

|∑

aij〈AiBjx, x〉|.

But the latter norm is familiar: Since dim(H) < ∞ and Ai, Bj are commuting self-adjoint unitaries,by Theorem 12.12 we have

supx∈BH ,Ai,Bj

|∑

aij〈AiBjx, x〉| = ‖a‖min = ‖a‖ℓ1⊗H′ℓ1 ,

so thatQM(a)max = ‖a‖min = ‖a‖ℓ1⊗H′ℓ1 .

Thus, if we set ‖a‖∨ = ‖a‖ℓn1

∨⊗ℓn

1

, then GT (see Theorem 1.15) says

‖a‖∨ ≤ ‖a‖min ≤ KRG ‖a‖∨

that is precisely equivalent to:

HV (a)max ≤ QM(a)max ≤ KRG HV (a)max.

But the covariance ξij can be physically measured, and hence also |∑ aijξij | for a fixed suitablechoice of a, so one can obtain an experimental answer for the maximum over all choices of (ξij)

EXP (a)max

and (for well chosen a) it deviates from the HV value. Indeed, if a is the identity matrix there isno deviation since we know Ai = −Bi, and hence

∑aijξij = −n, but for a non trivial choice of a,

for instance for

a =

(1 1−1 1

),

a deviation is found, indeed a simple calculation shows that for this particular choice of a:

QM(a)max =√

2 HV (a)max.

In fact the experimental data strongly confirms the QM predictions:

HV (a)max < EXP (a)max ≃ QM(a)max.

GT then appears as giving a bound for the deviation:

HV (a)max < QM(a)max but QM(a)max ≤ KRG HV (a)max.

At this point it is quite natural to wonder, as Tsirelson did in [146],what happens in caseof three observers A,B, and C (Charlie !), or even more. One wonders whether the analogousdeviation is still bounded or not. This was recently answered by Marius Junge with Perez-Garcia,Wolf, Palazuelos, Villanueva [34]. The analogous question for three separated observers A, B, Cbecomes: Given

a =∑

aijkei ⊗ ej ⊗ ek ∈ ℓn1 ⊗ ℓn

1 ⊗ ℓn1 ⊂ C∗(Fn) ⊗min C∗(Fn) ⊗min C∗(Fn).

65

Is there a constant K such that‖a‖min ≤ K‖a‖∨?

The answer is No! In fact, they get on ℓ2n2

1 ⊗ ℓ1 ⊗ ℓ1

K ≥ c√

n,

with c > 0 independent of n. We give a complete proof of this in the next section.For more information on the connection between Grothendieck’s and Bell’s inequalities, see

[74, 1, 31, 17, 62, 56].

20. Trilinear counterexamples

Many researchers have looked for a trilinear version of GT. The results have been mostly negative(see however [12, 13]). Note that a bounded trilinear form on C(S1)×C(S2)×C(S3) is equivalentto a bounded linear map C(S1) → (C(S2)⊗C(S3))

∗ and in general the latter will not factor througha Hilbert space. For a quick way to see this fact from the folklore, observe that ℓ2 is a quotientof L∞ and L∞ ≃ C(S) for some S, so that ℓ2⊗ℓ2 is a quotient of C(S)⊗C(S) and hence B(ℓ2)embeds into (C(S)⊗C(S))∗. But c0 or ℓ∞ embeds into B(ℓ2) and obviously this embedding doesnot factor through a Hilbert space (or even through any space not containing c0!). However, withthe appearance of the operator space versions, particularly because of the very nice behaviour ofthe Haagerup tensor product (see §14), several conjectural trilinear versions of GT reappeared andseemed quite plausible (see also §21 below).

If A, B are commutative C∗-algebras it is very easy to see that

‖t‖h = ‖t‖H !

Therefore, the classical GT says that, for all t in A⊗B, ‖t‖∧ ≤ KG‖t‖h (see (1.15) or Theorem 2.4above). Equivalently for all t in A∗ ⊗ B∗ or for all t in L1(µ) ⊗ L1(µ

′) (see Th. 1.5)

‖t‖H′ = ‖t‖h′ ≤ KG‖t‖∨.

But then a sort of miracle happens: ‖ · ‖h is self-dual, i.e. whenever t ∈ E∗ ⊗ F ∗ (E, F arbitraryo.s.) we have (see Remark 14.5)

‖t‖h′ = ‖t‖h.

Note however that here ‖t‖h means ‖t‖E∗⊗hF ∗ where E∗, F ∗ must be equipped with the dual o.s.s.described in §11. In any case, if A∗, B∗ are equipped with their dual o.s.s. we find ‖t‖h ≤ KG‖t‖∨and a fortiori ‖t‖min ≤ KG‖t‖∨, for any t in A∗ ⊗ B∗. In particular, let A = B = c0 so thatA∗ = B∗ = ℓ1 (equipped with its dual o.s.s.). We obtain

Corollary 20.1. For any t in ℓ1 ⊗ ℓ1

(20.1) ‖t‖min ≤ KG‖t‖∨.

Recall here (see Remark 13.13) that ℓ1 with its maximal o.s.s. (dual to the natural one for c0)can be “realized” as an operator space inside C∗(F∞): One simply maps the n-th basis vector en

to the unitary Un associated to the n-th free generator.It then becomes quite natural to wonder whether the trilinear version of Corollary 20.1 holds.

This was disproved by a rather delicate counterexample due to Junge (unpublished), a new versionof which was later published in [34].

66

Theorem 20.2 (Junge). The injective norm and the minimal norm are not equivalent on ℓ1⊗ℓ1⊗ℓ1.

Lemma 20.3. Let εij | 1 ≤ i, j ≤ n be an i.i.d. family representing n2 independent choices ofsigns εij = ±1. Let Rn ⊂ L1(Ω,A, P) be their linear span and let wn : Rn → Mn be the linearmapping taking εij to eij. We equip Rn with the o.s.s. induced on it by the maximal o.s.s. onL1(Ω,A, P) (see Remark 13.13). Then ‖wn‖cb ≤ 31/2n1/2.

Proof. Consider a function Φ ∈ L∞(P; Mn) (depending only on (εij) 1 ≤ i, j ≤ n) such that∫Φεij = eij . Let wn : L1 → Mn be the operator defined by wn(x) =

∫xΦ dP, so that wn(εij) = eij .

Then wn extends wn so that‖wn‖cb ≤ ‖wn‖cb

but by (13.4) we have‖wn‖cb = ‖Φ‖L∞(P)⊗minMn

= ‖Φ‖L∞(P);Mn).

But now by (9.10)

inf

‖Φ‖∞

∣∣∣∫

Φεij = eij

≤ 31/2‖(eij)‖RC

and it is very easy to check that

‖(eij)‖R = ‖(eij)‖C = n1/2

so we conclude that ‖wn‖cb ≤ 31/2n1/2.

Remark 20.4. Let Gn ⊂ L1 denote the linear span of a family of n2 independent standard complexGaussian random variables gC

ij. Let Wn : Gn → Mn be the linear map taking gCij to eij . An

identical argument to the preceding one shows that ‖Wn‖cb ≤ 21/2n1/2.

Proof of Theorem 20.2. We will define below an element T ∈ Rn ⊗RN ⊗RN ⊂ L1 ⊗L1 ⊗L1, andwe note immediately that (by injectivity of ‖ · ‖min) we have

(20.2) ‖T‖Rn⊗minRN⊗minRN= ‖T‖L1⊗minL1⊗minL1 .

Consider the natural isometric isomorphism

ϕN : MN ⊗min MN −→ (ℓN2 ⊗2 ℓN

2 )∨⊗ (ℓN

2 ⊗2 ℓN2 )

defined byϕN (epq ⊗ ers) = (ep ⊗ er) ⊗ (eq ⊗ es).

Let Y ′i , Y ′′

j be the matrices in MN appearing in Corollary 16.7. Let χN : (ℓN2 ⊗2ℓ

N2 )

∨⊗ (ℓN

2 ⊗2ℓN2 ) −→

RN

∨⊗ RN be the linear isomorphism defined by

χN ((ep ⊗ er) ⊗ (eq ⊗ es)) = εpr ⊗ εqs.

Since

(20.3)∥∥∥∑

αprεpr

∥∥∥1≤

∥∥∥∑

αprεpr

∥∥∥2

=(∑

|αpr|2)1/2

for all scalars (αpr), we have ‖χN‖ ≤ 1 and hence ‖χNϕN : MN ⊗min MN → RN

∨⊗ RN‖ ≤ 1. We

define T ∈ Rn ⊗RN ⊗RN by

T =∑

εij ⊗ (χNϕN )(Y ′i ⊗ Y ′′

j ).

67

By Corollary 16.7 we can choose Y ′i , Y ′′

j so that (using (20.3) on the first coordinate)

(20.4) ‖T‖Rn

∨⊗RN

∨⊗RN

≤ 4 + ε.

But by the preceding Lemma and by (13.7), we have

(20.5) ‖(wn ⊗ wN ⊗ wN )(T )‖Mn⊗minMN⊗minMN≤ 33/2n1/2N‖T‖Rn⊗minRN⊗minRN

.

Then(wN ⊗ wN )(χNϕN )(epq ⊗ ers) = epr ⊗ eqs

so that if we rewrite (wn ⊗ wN ⊗ wN )(T ) ∈ Mn ⊗min MN ⊗min MN as an element T of

(ℓn2 ⊗2 ℓN

2 ⊗2 ℓN2 )∗

∨⊗ (ℓn

2 ⊗2 ℓN2 ⊗2 ℓN

2 ) then we find

T =(∑

ipqei ⊗ ep ⊗ eqY

′i (p, q)

)⊗

(∑jrs

ej ⊗ er ⊗ esY′′j (r, s)

).

Note that since T is of rank 1

‖T‖∨ =(∑

ipq|Y ′

i (p, q)|2)1/2 (∑

jrs|Y ′′

j (r, s)|2)1/2

and hence by (16.16)

‖(wn ⊗ wN ⊗ wN )(T )‖Mn⊗minMN⊗minMN= ‖T‖∨ ≥ (1 − ε)nN.

Thus we conclude by (20.4) and (20.5) that

‖T‖L1

∨⊗L1

∨⊗L1

≤ 4 + ε but ‖T‖L1⊗minL1⊗minL1 ≥ 3−3/2(1 − ε)n1/2.

This completes the proof with L1 instead of ℓ1. Note that we obtain a tensor T in

ℓ2n2

1 ⊗ ℓ2N2

1 ⊗ ℓ2N2

1 .

Remark 20.5. Using [125, p. 16], one can reduce the dimension 2n2to one ≃ n4 in the preceding

example.

Remark 20.6. The preceding construction establishes the following fact: there is a constant c > 0such that for any positive integer n there is a finite rank map u : ℓ∞ → ℓ1 ⊗min ℓ1, associatedto a tensor t ∈ ℓ1 ⊗ ℓ1 ⊗ ℓ1, such that the map un : Mn(ℓ∞) → Mn(ℓ1 ⊗min ℓ1) defined byun([aij ]) = [u(aij)] satisfies

(20.6) ‖un‖ ≥ c√

n‖u‖.

A fortiori of course ‖u‖cb = supn ‖un‖ ≥ c√

n‖u‖. Note that by (13.9) ‖u‖cb = ‖t‖ℓ1⊗minℓ1⊗minℓ1 andby GT (see (20.1)) we have ‖t‖

ℓ1∨⊗ℓ1

∨⊗ℓ1

≤ ‖u‖ = ‖t‖ℓ1

∨⊗(ℓ1⊗minℓ1)

≤ KCG‖t‖ℓ1

∨⊗ℓ1

∨⊗ℓ1

.

In this form (as observed in [34]) the estimate (20.6) is sharp. Indeed, it is rather easy to showthat, for some constant c′, we have ‖un‖ ≤ c′

√n‖u‖ for any operator space E and any map

u : ℓ∞ → E (this follows from the fact that the identity of Mn admits a factorization through ann2-dimensional quotient Q of L∞ of the form IMn

= ab where b : Mn → Q and a : Q → Mn satisfy‖b‖‖a‖cb ≤ c′

√n).

68

Remark. In [34], there is a different proof of (16.15) that uses the fact that the von Neumann algebraof F∞ embeds into an ultraproduct (von Neumann sense) of matrix algebras. In this approach,one can use, instead of random matrices, the residual finiteness of free groups. This leads to thefollowing substitute for Corollary 16.7: Fix n and ε > 0, then for all N ≥ N0(n, ε) there are N ×Nmatrices Y ′

i , Y ′′j and ξij such that Y ′

i , Y ′′j are unitary (permutation matrices) satisfying:

∀(αij) ∈ Cn2

∥∥∥∑n

i,j=1αij(Y

′i ⊗ Y ′′

j + ξij)∥∥∥

MN2

≤ (4 + ε)(∑

|αij |2)1/2

,

and

∀i, j = 1, . . . , n N−1 tr(ξ2ij) < ε.

This has the advantage of a deterministic choice of Y ′i , Y ′′

j but the inconvenient that (ξij) is notexplicit, so only the “bulk of the spectrum” is controlled. An entirely explicit and non-randomexample proving Theorem 20.2 is apparently still unknown.

21. Some open problems

In Theorem 18.1, we obtain an equivalent norm on CB(A, B∗), and we can claim that this elucidatesthe Banach space structure of CB(A, B∗) in the same way as GT does when A, B are commutativeC∗-algebras (see Theorem 2.1). But since the space CB(A, B∗) comes naturally equipped with ano.s.s. (see Remark 13.8) it is natural to wonder whether we can actually improve Theorem 18.1 toalso describe the o.s.s. of CB(A, B∗). In this spirit, the natural conjecture (formulated by DavidBlecher in the unpublished problem book of an October 1993 conference at Texas A& M University)is as follows:

Problem 21.1. Consider a jointly c.b. bilinear form ϕ : A × B → B(H) (this means that ϕ ∈CB(A, CB(B, B(H))) or equivalently that ϕ ∈ CB(B, CB(A, B(H))) with the obvious abuse).

Is it true that ϕ can be decomposed as ϕ = ϕ1 + tϕ2 where ϕ1 ∈ CB(A ⊗h B, B(H)) andϕ2 ∈ CB(B ⊗h A, B(H)) where tϕ2(a ⊗ b) = ϕ2(b ⊗ a)?

Of course (as can be checked using Remark 14.3, suitably generalized) the converse holds: anymap ϕ with such a decomposition is c.b.

As far as we know this problem is open even in the commutative case, i.e. if A, B are bothcommutative. The problem clearly reduces to the case B(H) = Mn with a constant C independentof n such that inf‖ϕ1‖cb + ‖ϕ2‖cb ≤ C‖ϕ‖cb. In this form (with an absolute constant C), thecase when A, B are matrix algebras is also open. Apparently, this might even be true when A, Bare exact operator spaces, in the style of Theorem 17.2.

In a quite different direction, the trilinear counterexample in §20 does not rule out a certaintrilinear version of the o.s. version of GT that we will now describe:

Let A1, A2, A3 be three C∗-algebras. Consider a (jointly) c.b. trilinear form ϕ : A1×A2×A3 →C, with c.b. norm ≤ 1. This means that for any n ≥ 1 ϕ defines a trilinear map of norm ≤ 1

ϕ[n] : Mn(A1) × Mn(A2) × Mn(A3) −→ Mn ⊗min Mn ⊗min Mn ≃ Mn3 .

Equivalently, this means that ϕ ∈ CB(A1, CB(A2, A∗3)) or more generally this is the same as

ϕ ∈ CB(Aσ(1), CB(Aσ(2), A∗σ(3)))

for any permutation σ of 1, 2, 3, and the obvious extension of Proposition 13.9 is valid.

69

Problem 21.2. Let ϕ : A1 × A2 × A3 → C be a (jointly) c.b. trilinear form as above. Is it truethat ϕ can be decomposed as a sum

ϕ =∑

σ∈S3

ϕσ

indexed by the permutations σ of 1, 2, 3, such that for each such σ, there is a

ψσ ∈ CB(Aσ(1) ⊗h Aσ(2) ⊗h Aσ(3), C)

such thatϕσ(a1 ⊗ a2 ⊗ a3) = ψσ(aσ(1) ⊗ aσ(2) ⊗ aσ(3)) ?

Note that such a statement would obviously imply the o.s. version of GT given as Theorem 18.1as a special case.

Remark. See [13, p.104] for an open question of the same flavor as the preceding one, for boundedtrilinear forms on commutative C∗-algebras, but involving factorizations through Lp1 × Lp2 × Lp3

with (p1, p2, p3) such that p−11 + p−1

2 + p−13 = 1 varying depending on the trilinear form ϕ.

22. GT in graph theory and computer science

In [3], the Grothendieck constant of a (finite) graph G = (V, E) is introduced, as the smallestconstant K such that, for every a : E → R, we have

(22.1) supf : V →S

∑

s,t∈E

a(s, t)〈f(s), f(t)〉 ≤ K supf : V →−1,1

∑

s,t∈E

a(s, t)f(s)f(t)

where S is the unit sphere of H = ℓ2. Note that we may replace H by the span of the range off and hence we may always assume dim(H) ≤ |V |. We will denote by K(G) the smallest such K.Consider for instance the complete bipartite graph CBn on vertices V = In∪Jn with In = 1, . . . , n,Jn = n + 1, . . . , 2n with (i, j) ∈ E ⇔ i ∈ In, j ∈ Jn. In that case (22.1) reduces to (2.5) and wehave

(22.2) K(CBn) = KRG(n) and sup

n≥1K(CBn) = KR

G.

If G = (V ′, E′) is a subgraph of G (i.e. V ′ ⊂ V and E′ ⊂ E) then obviously

K(G′) ≤ K(G).

Therefore, for any bipartite graph G we have

K(G) ≤ KRG.

However, this constant does not remain bounded for general (non-bipartite) graphs. In fact, it isknown (cf. [100] and [99] independently) that there is an absolute constant C such that for any Gwith no selfloops (i.e. (s, t) /∈ E when s = t)

(22.3) K(G) ≤ C(log(|V |) + 1).

Moreover by [3] this logarithmic growth is asymptotically optimal, thus improving Kashin andSzarek’s lower bound [72] that answered a question raised by Megretski [99] (see the above Remark2.12). Note however that the log(n) lower bound found in [3] for the complete graph on n vertices

70

was somewhat non-constructive, and more recently, explicit examples were produced in [5] butyielding only a log(n)/ log log(n) lower bound.

Let us now describe briefly the motivation behind these notions. One major breakthrough wasGoemans and Williamson’s paper [35] that uses semidefinite programming (ignorant readers likethe author will find [91] very illuminating). The origin can be traced back to a famous problemcalled MAX CUT. By definition, a cut in a graph G = (V, E) is a set of edges connecting a subsetS ⊂ V of the vertices to the complementary subset V − S. The MAX CUT problem is to find acut with maximum cardinality.

MAX CUT is known to be hard in general, in precise technical terms it is NP-hard. Recall thatthis implies that P = NP would follow if it could be solved in polynomial time. Alon and Naor [4]proposed to compare MAX CUT to another problem that they called the CUT NORM problem:We are given a real matrix (aij)i∈R,j∈C we want to compute efficiently

Q = maxI⊂R,J⊂C

∣∣∣∣∣∣

∑

i∈I,j∈J

aij

∣∣∣∣∣∣,

and to find a pair (I, J) realizing the maximum. Of course the connection to GT is that thisquantity Q is such that

Q ≤ Q′ ≤ 4Q where Q′ = supxi,yj∈−1,1

∑aijxiyj .

So roughly computing Q is reduced to computing Q′, and finding (I, J) to finding a pair of choicesof signs. Recall that S denotes the unit sphere in Hilbert space. Then precisely Grothendieck’sinequality (2.5) tells us that

1

KGQ′′ ≤ Q′ ≤ Q′′ where Q′′ = sup

xi,yj∈S

∑aij〈xi, yj〉.

The point is that computing Q′ in polynomial time is not known and very unlikely to becomeknown (in fact it would imply P = NP ) while the problem of computing Q′′ turns out to befeasible: Indeed, it falls into the category of semidefinite programming problems and these areknown to be solvable, within an additive error of ε, in polynomial time (in the length of theinput and in the logarithm of 1/ε), because one can exploit Hilbert space geometry, namely “theellipsoid method” (see [39, 40]). Inspired by earlier work by Lovasz-Schrijver and Alizadeh (see[35]), Goemans and Williamson introduced, in the context of MAX CUT, a geometric method toanalyze the connection of a problem such as Q′′ with the original combinatorial problem such asQ′. In this context, Q′′ is called a “relaxation” of the corresponding problem Q′. Grothendieck’sinequality furnishes a new way to approach harder problems such as the CUT NORM problem Q′′.The CUT NORM problem is indeed harder than MAX CUT since Alon and Naor [4] showed thatMAX CUT can be cast a special case of CUT NORM; this implies in particular that CUT NORMis also NP hard.It should be noted that by known results: There exists ρ < 1 such that even being able to computeQ′ up to a factor > ρ in polynomial time would imply P = NP . The analogue of this fact (withρ = 16/17) for MAX CUT goes back to Hastad [55] (see also [105]). Actually, according to[130, 131], for any 0 < K < KG, assuming a strengthening of P 6= NP called the “unique gamesconjecture”, it is NP-hard to compute any quantity q such that K−1q ≤ Q′. While, for K > KG,we can take q = Q′′ and then compute a solution in polynomial time by semidefinite programming.So in this framework KG seems connected to the P = NP problem !

71

But the CUT NORM (or MAX CUT) problem is not just a maximization problem as the precedingones of computing Q′ or Q′′. There one wants to find the vectors with entries ±1 that achieve themaximum or realize a value greater than a fixed factor ρ times the maximum. The semidefiniteprogramming produces 2n vectors in the n-dimensional Euclidean sphere, so there is a delicate“rounding problem” to obtain instead vectors in −1, 1n. In [4], the authors transform a knownproof of GT into one that solves efficiently this rounding problem. They obtain a deterministicpolynomial time algorithm that produces x, y ∈ −1, 1n such that

∑aij〈xi, yj〉 ≥ 0.03 Q′′ (and a

fortiori ≥ 0.03 Q′). Let ρ = 2 log(1 +√

2)/π > 0.56 be the inverse of Krivine’s upper bound forKG. They also obtain a randomized algorithm for that value of ρ. Here randomized means thatthe integral vectors are random vectors so that the expectation of

∑aij〈xi, yj〉 is ≥ ρ Q′′.

The papers [3, 4] ignited a lot of interest in the computer science literature. Here are a fewsamples: In [2] the Grothendieck constant of the random graph on n vertices is shown to be oforder log(n) (with probability tending to 1 as n → ∞), answering a question left open in [3]. In[76], the “hardness” of computing the norm of a matrix on ℓn

p × ℓnp is evaluated depending on the

value of 2 ≤ p ≤ ∞. The Grothendieck related case is of course p = ∞. In [20], the authors studya generalization of MAX CUT that they call MAXQP. In [130, 131], a (programmatic) approachis proposed to compute KG efficiently. More references are [2, 145, 90, 130, 131, 75, 76].

Acknowledgment. This paper is partially based on the author’s notes for a minicourse at IIScBangalore (10th discussion meeting in Harmonic Analysis, Dec. 2007) and a lecture at IHES(Colloque en l’honneur d’Alexander Grothendieck, Jan. 2009). I thank the organizers for theresulting stimulation. I am very grateful to Mikael de la Salle for useful conversations, and forvarious improvements. Thanks are also due to Maurice Rojas, Assaf Naor, Hermann Konig, KateJuschenko for useful correspondence.

References

[1] A. Acın, N. Gisin and B. Toner, Grothendieck’s constant and local models for noisy entangledquantum states, Phys. Rev. A (3) 73 (2006), no. 6, part A, 062105, 5 pp.

[2] N. Alon and E. Berger, The Grothendieck constant of random and pseudo-random graphs,Discrete Optim. 5 (2008), no. 2, 323–327.

[3] N. Alon, K. Makarychev, Y. Makarychev, and A. Naor, Quadratic forms on graphs, Invent.Math. 163 (2006), no. 3, 499–522.

[4] N. Alon and A. Naor, Approximating the cut-norm via Grothendieck’s inequality, SIAM J.Comput. 35 (2006), no. 4, 787–803 (electronic).

[5] S. Arora, E. Berger, E. Hazan, G. Kindler, and M. Safra, On non-approximability forquadratic programs, Preprint, to appear.

[6] A. Aspect, Bell’s theorem: the naive view of an experimentalist, Quantum [Un]speakables(Vienna, 2000), 119–153, Springer, Berlin, 2002.

[7] A. Aspect, Testing Bell’s inequalities, Quantum Reflections, 69–78, Cambridge Univ. Press,Cambridge, 2000.

[8] J. Audretsch, Entangled Systems, WILEY-VCH, VerlagGmbH & Co., KGaA, Weinheim, 2007.

72

[9] J. Bergh and J. Lofstrom, Interpolation Spaces. An Introduction, Springer Verlag, New York,1976.

[10] D. Blecher, Tracially completely bounded multilinear maps on C∗-algebras, J. London Math.Soc. 39 (1989), no. 3, 514–524.

[11] D. Blecher, Generalizing Grothendieck’s program, “Function spaces”, K. Jarosz (ed.), LectureNotes in Pure and Applied Math., Vol. 136, Marcel Dekker, 1992.

[12] R.C. Blei, Multidimensional extensions of the Grothendieck inequality and applications, Ark.Mat. 17 (1979), 51–68.

[13] R.C. Blei, Analysis in Integer and Fractional Dimensions, Cambridge University Press, Cam-bridge, 2001.

[14] M. Bozejko and G. Fendler, Herz–Schur multipliers and completely bounded multipliers of theFourier algebra of a locally compact group, Boll. Unione Mat. Ital. (6) 3-A (1984), 297–302.

[15] J. Briet, F.M. de Oliveira Filho and F. Vallentin, The positive semidefinite Grothendieckproblem with rank constraint, pp. 31-42 in Proceedings of the 37th International Colloquiumon Automata, Languages and Programming, ICALP 2010 S. Abramsky, et. al. (eds.), Part I,LNCS 6198, 2010.

[16] J. Briet, F.M. de Oliveira Filho and F. Vallentin, Grothendieck inequalities for semidefiniteprograms with rank constraint, arXiv:1011.1754v1 [math.OC]

[17] J. Briet, H. Burhman and B. Toner, A generalized Grothendieck inequality and entanglementin XOR games, arXiv:0901.2009v1[quant-ph].

[18] N.P. Brown and N. Ozawa, C∗-algebras and finite-dimensional approximations, GraduateStudies in Mathematics, 88, American Mathematical Society, Providence, RI, 2008.

[19] A. Buchholz, Optimal constants in Khintchine type inequalities for Fermions, Rademachersand q-Gaussian operators, Bull. Polish Acad. Sci. Math. 53 (2005), 315–321.

[20] M. Charikar and A. Wirth, Maximizing quadratic programs: extending Grothendieck’s in-equality, FOCS (2004), 54–60.

[21] M.D. Choi and E. Effros, Nuclear C∗-algebras and injectivity: the general case, Indiana Univ.Math. J. 26 (1977), 443–446.

[22] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. Math. (2)104 (1976), 73–115.

[23] A.M. Davie, Matrix norms related to Grothendieck’s inequality, Banach spaces (Columbia,MO, 1984), 22–26, Lecture Notes in Math., 1166, Springer, Berlin, 1985.

[24] J. Diestel, J.H. Fourie and J. Swart, The metric theory of tensor products. Grothendieck’sresume revisited, American Mathematical Society, Providence, RI, 2008.

[25] J. Diestel, H. Jarchow and A. Pietsch, Operator ideals, Handbook of the Geometry of BanachSpaces, Vol. I, 437–496, North-Holland, Amsterdam, 2001.

73

[26] J. Dixmier, Les anneaux d’operateurs de classe finie, Ann. Sci. Ecole Norm. Sup. 66 (1949),209–261.

[27] E. Dubinsky, A. PeÃlczynski and H.P. Rosenthal, On Banach spaces X for which Π2(L∞, X) =B(L∞, X), Studia Math. 44 (1972), 617–648.

[28] K. Dykema and K. Juschenko, Matrices of unitary moments, Preprint (arXiv:0901.0288).

[29] E.G. Effros and Z.J. Ruan, A new approach to operator spaces, Canadian Math. Bull. 34(1991), 329–337.

[30] E.G. Effros and Z.J. Ruan, Operator Spaces, The Clarendon Press, Oxford University Press,New York, 2000, xvi+363 pp.

[31] P.C. Fishburn and J.A. Reeds, Bell inequalities, Grothendieck’s constant, and root two, SIAMJ. Discrete Math. 7 (1994), no. 1, 48–56.

[32] T. Fritz, Tsirelson’s problem and Kirchberg’s conjecture, 2010 (arXiv:1008.1168).

[33] T.W. Gamelin and S.V. Kislyakov, Uniform algebras as Banach spaces, Handbook of theGeometry of Banach Spaces, Vol. I, 671–706, North-Holland, Amsterdam, 2001.

[34] D. Perez-Garcıa, M.M. Wolf, C. Palazuelos, I. Villanueva, and M. Junge, Unbounded violationof tripartite Bell inequalities, Comm. Math. Phys. 279 (2008), no. 2, 455–486.

[35] M.X. Goemans and D.P. Williamson, Improved approximation algorithms for maximum cutand satisfiability problems using semidefinite programming, J. Assoc. Comput. Mach. 42(1995), 1115–1145.

[36] A. Grothendieck, Resume de la theorie metrique des produits tensoriels topologiques, Boll.Soc. Mat. Sao-Paulo 8 (1953), 1–79. Reprinted in Resenhas 2 (1996), no. 4, 401–480.

[37] A. Grothendieck, Sur certaines classes de suites dans les espaces de Banach et le theoremede Dvoretzky–Rogers, Bol. Soc. Mat. Sao Paulo 8 (1953), 81–110. Reprinted in Resenhas 3(1998), no. 4, 447–477.

[38] A. Grothendieck, Produits tensoriels topologiques et espaces nucleaires (French), Mem. Amer.Math. Soc. 1955 (1955), no. 16, 140 pp.

[39] M. Grotschel, L. Lovasz and A. Schrijver, The ellipsoid method and its consequences incombinatorial optimization, Combinatorica 1 (1981), no. 2, 169–197.

[40] M. Grotschel, L. Lovasz and A. Schrijver, Geometric algorithms and combinatorial optimiza-tion, Second edition, Algorithms and Combinatorics, 2, Springer-Verlag, Berlin, 1993, xii+362pp.

[41] U. Haagerup, An example of a non-nuclear C∗-algebra which has the metric approximationproperty, Inventiones Mat. 50 (1979), 279–293.

[42] U. Haagerup, The best constants in the Khintchine inequality, Studia Math. 70 (1981), 231–283 (1982).

[43] U. Haagerup, Solution of the similarity problem for cyclic representations of C∗-algebras,Ann. Math. 118 (1983), 215–240.

74

[44] U. Haagerup, All nuclear C∗-algebras are amenable, Invent. Math. 74 (1983), no. 2, 305–319.

[45] U. Haagerup, The Grothendieck inequality for bilinear forms on C∗-algebras, Adv. Math. 56(1985), no. 2, 93–116.

[46] U. Haagerup, A new upper bound for the complex Grothendieck constant, Israel J. Math. 60(1987), no. 2, 199–224.

[47] U. Haagerup and G. Itoh, Grothendieck type norms for bilinear forms on C∗-algebras, J.Operator Theory 34 (1995), 263–283.

[48] U. Haagerup and M. Musat, On the best constants in noncommutative Khintchine-typeinequalities, J. Funct. Anal. 250 (2007), 588–624.

[49] U. Haagerup and M. Musat, The Effros–Ruan conjecture for bilinear forms on C∗-algebras,Invent. Math. 174 (2008), 139–163.

[50] U. Haagerup and G. Pisier, Factorization of analytic functions with values in non-commutativeL1-spaces, Canad. J. Math. 41 (1989), 882–906.

[51] U. Haagerup and G. Pisier, Linear operators between C∗-algebras, Duke Math. J. 71 (1993),889–925.

[52] U. Haagerup and S. Thorbjørnsen, Random matrices and K-theory for exact C∗-algebras,Doc. Math. 4 (1999), 341–450 (electronic).

[53] U. Haagerup and S. Thorbjørnsen, A new application of random matrices: Ext(C∗red(F2)) is

not a group, Ann. of Math. (2) 162 (2005), no. 2, 711–775.

[54] A. Harcharras, Fourier analysis, Schur multipliers on Sp and non-commutative Λ(p)-sets,Studia Math. 137 (1999), no. 3, 203–260.

[55] J. Hastad, Some optimal inapproximability results, J. ACM 48 (2001), 798–859.

[56] H. Heydari, Quantum correlation and Grothendieck’s constant, J. Phys. A 39 (2006), no. 38,11869–11875.

[57] K. Ito and M. Nisio, On the convergence of sums of independent Banach space valued randomvariables, Osaka J. Math. 5 (1968), 35–48.

[58] T. Itoh, On the completely bounded map of a C∗-algebra to its dual space, Bull. LondonMath. Soc. 19 (1987), 546–550.

[59] K. John, On the compact nonnuclear operator problem, Math. Ann. 287 (1990), 509–514.

[60] M. Junge, Embedding of the operator space OH and the logarithmic ‘little Grothendieckinequality’, Invent. Math. 161 (2005), no. 2, 225–286.

[61] M. Junge, M. Navascues, C. Palazuelos, D. Perez-Garcıa, V.B. Scholz, and R.F. Werner,Connes’ embedding problem and Tsirelson’s problem, 2010, Preprint (arXiv:1008.1142).

[62] M. Junge and C. Palazuelos, Large violations of Bell’s inequalities with low entanglement,(arXiv:1007.3043).

75

[63] M. Junge and J. Parcet, Rosenthal’s theorem for subspaces of noncommutative Lp, DukeMath. J. 141 (2008), no. 1, 75–122.

[64] M. Junge and J. Parcet, Maurey’s factorization theory for operator spaces, Math. Ann. 347(2010), no. 2, 299–338.

[65] M. Junge and J. Parcet, Mixed-norm inequalities and operator space Lp embedding theory,Mem. Amer. Math. Soc. 203 (2010), no. 953, vi+155 pp.

[66] M. Junge and G. Pisier, Bilinear forms on exact operator spaces and B(H) ⊗ B(H), Geom.Funct. Anal. 5 (1995), no. 2, 329–363.

[67] M. Junge, and Q. Xu, Representation of certain homogeneous Hilbertian operator spaces andapplications, Invent. Math. 179 (2010), no. 1, 75–118,

[68] R. Kadison and J. Ringrose, Fundamentals of the Theory of Operator Algebras, Vol. II, Ad-vanced Theory, Corrected reprint of the 1986 original. Graduate Studies in Mathematics, 16,American Mathematical Society, Providence, RI, 1997.

[69] J.P. Kahane, Some random series of functions, Second edition, Cambridge Studies in AdvancedMathematics, 5, Cambridge University Press, Cambridge, 1985.

[70] S. Kaijser and A.M. Sinclair, Projective tensor products of C∗-algebras, Math. Scand. 55(1984), no. 2, 161–187.

[71] N.J. Kalton, Rademacher series and decoupling, New York J. Math. 11 (2005), 563–595(electronic).

[72] B. Kashin and S. Szarek, The Knaster problem and the geometry of high-dimensional cubes,C.R. Math. Acad. Sci. Paris 336 (2003), 931–936.

[73] B. Kashin and S. Szarek, On the Gram matrices of systems of uniformly bounded functions(Russian), Tr. Mat. Inst. Steklova 243 (2003), Funkts. Prostran., Priblizh., Differ. Uravn.,237–243; translation in Proc. Steklov Inst. Math. 2003, no. 4 (243), 227–233.

[74] L.A. Khalfin and B.S. Tsirelson, Quantum and quasiclassical analogs of Bell inequalities,Symposium on the Foundations of Modern Physics (Joensuu, 1985), 441–460, World Sci.Publishing, Singapore, 1985.

[75] S. Khot and A. Naor, Sharp kernel clustering algorithms and their associated Grothendieckinequalities.

[76] G. Kindler, A. Naor and G. Schechtman, The UGC hardness threshold of the Lp Grothendieckproblem, Math. Oper. Res. 35 (2010), no. 2, 267–283.

[77] E. Kirchberg, On nonsemisplit extensions, tensor products and exactness of group C∗-algebras, Invent. Math. 112 (1993), 449–489.

[78] S.V. Kislyakov, Linear and complex analysis problem book, Section 6.5, Lecture Notes inMath., 1043, Springer, Berlin, 1984.

[79] S.V. Kislyakov, Absolutely summing operators on the disc algebra (Russian), Algebra i Analiz3 (1991), no. 4, 1–77; translation in St. Petersburg Math. J. 3 (1992), 705–774.

76

[80] S.V. Kislyakov, Banach spaces and classical harmonic analysis, Handbook of the Geometry ofBanach Spaces, Vol. I, 871–898, North-Holland, Amsterdam, 2001.

[81] H. Konig, On the complex Grothendieck constant in the n-dimensional case, Geometry ofBanach Spaces (Strobl, 1989), 181–198, London Math. Soc. Lecture Note Ser., 158, CambridgeUniv. Press, Cambridge, 1990.

[82] H. Konig, On an extremal problem originating in questions of unconditional convergence, inRecent Progress in Multivariate Approximation, Conf. Bommerholz 2000, Birkhauser, 185–192, 2001.

[83] J.L. Krivine, Theoremes de factorisation dans les espaces reticules, Seminaire Maurey–Schwartz 1973–1974: Espaces Lp, applications radonifiantes et geometrie des espaces de Ba-nach, Exp. Nos. 22 et 23, 22 pp. Centre de Math., Ecole Polytech., Paris, 1974.

[84] J.L. Krivine, Sur la constante de Grothendieck, C.R. Acad. Sci. Paris Ser. A 284 (1977),445–446.

[85] J.L. Krivine, Constantes de Grothendieck et fonctions de type positif sur les spheres, Adv. inMath. 31 (1979), 16–30.

[86] S. Kwapien, On operators factorizable through Lp-space, Bull. Soc. Math. France Memoire31-32 (1972), 215–225.

[87] M. Ledoux, The Concentration of Measure Phenomenon, American Mathematical Society,Providence, RI, 2001.

[88] D.H. Leung, Factoring operators through Hilbert space, Israel J. Math. 71 (1990), 225–227.

[89] J. Lindenstrauss and A. PeÃlczynski, Absolutely summing operators in Lp-spaces and theirapplications, Studia Math. 29 (1968), 275–326.

[90] N. Linial and A. Shraibman, Lower bounds in communication complexity based on factor-ization norms, in Proceedings of the 39th Annual ACM Symposium on Theory of Computing,699–708, ACM, New York, 2007.

[91] L. Lovasz, Semidefinite programs and combinatorial optimization, Lecture Notes, MicrosoftResearch Redmont, WA98052

[92] F. Lust-Piquard, Inegalites de Khintchine dans Cp (1 < p < ∞), (French) [Khinchine in-equalities in Cp (1 < p < ∞)], C.R. Acad. Sci. Paris Ser. I Math. 303 (1986), no. 7, 289–292.

[93] F. Lust-Piquard, A Grothendieck factorization theorem on 2-convex Schatten spaces, IsraelJ. Math. 79 (1992), no. 2-3, 331–365.

[94] F. Lust-Piquard and G. Pisier, Noncommutative Khintchine and Paley inequalities, Ark. Mat.29 (1991), no. 2, 241–260.

[95] F. Lust-Piquard and Q. Xu, The little Grothendieck theorem and Khintchine inequalities forsymmetric spaces of measurable operators, J. Funct. Anal. 244 (2007), no. 2, 488–503.

[96] M. Braverman, K. Makarychev, Y. Makarychev, and A. Naor, On maximizers of oscillatoryGaussian kernels, paper in preparation.

77

[97] B. Maurey, Une nouvelle demonstration d’un theoreme de Grothendieck, Seminaire Maurey–Schwartz Annee 1972–1973: Espaces Lp et applications radonifiantes, Exp. No. 22, 7 pp.Centre de Math., Ecole Polytech., Paris, 1973.

[98] B. Maurey, Theoremes de factorisation pour les operateurs lineaires a valeurs dans les espacesLp, Asterisque, No. 11. Societe Mathematique de France, Paris, 1974, ii+163 pp.

[99] A. Megretski, Relaxations of quadratic programs in operator theory and system analysis(English summary), Systems, Approximation, Singular Integral Operators, and Related Topics(Bordeaux, 2000), 365–392, Oper. Theory Adv. Appl., 129, Birkhauser, Basel, 2001.

[100] A.R.C. Nemirovski and T. Terlaky, On maximization of quadratic form over intersection ofellipsoids with common center, Math. Programming 86 (1999), 463–473.

[101] T. Oikhberg, Direct sums of operator spaces, J. London Math. Soc. 64 (2001), 144–160.

[102] T. Oikhberg and G. Pisier, The “maximal” tensor product of operator spaces, Proc. EdinburghMath. Soc. 42 (1999), 267–284.

[103] A.M. Olevskii, Fourier series with respect to general orthogonal systems, Translated fromthe Russian by B.P. Marshall and H.J. Christoffers, Ergebnisse der Mathematik und ihrerGrenzgebiete, Band 86, Springer-Verlag, New York-Heidelberg, 1975.

[104] N. Ozawa, About the QWEP conjecture, Internat. J. Math. 15 (2004), no. 5, 501–530.

[105] C. Papadimitriou, and M. Yannakakis, Optimization, approximation, and complexity classes,J. Comput. System Sci. 43 (1991), 425–440.

[106] V. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in AdvancedMathematics, 78, Cambridge University Press, Cambridge, 2002.

[107] V. Paulsen and M. Raghupathi, Representations of logmodular algebras, to appear in Trans.Amer. Math. Soc.

[108] A. Peralta, Little Grothendieck’s theorem for real JB∗-triples, Math. Z. 237 (2001), 531–545.

[109] A. Peralta, New advances on the Grothendieck’s inequality problem for bilinear forms onJB*-triples, Math. Inequal. Appl. 8 (2005), 7–21.

[110] A. Peres, Quantum theory: concepts and methods, Fundamental Theories of Physics, 57,Kluwer Academic Publishers Group, Dordrecht, 1993.

[111] A. Pietsch, Operator Ideals, North-Holland Amsterdam, 1978.

[112] G. Pisier, Grothendieck’s theorem for noncommutative C∗-algebras, With an appendix onGrothendieck’s constants, J. Funct. Anal. 29 (1978), no. 3, 397–415.

[113] G. Pisier, Probabilistic methods in the geometry of Banach spaces, Probability and Analysis(Varenna, 1985), 167–241, Lecture Notes in Math., 1206, Springer, Berlin, 1986.

[114] G. Pisier, Factorization of operators through Lp∞ or Lp1 and noncommutative generalizations,Math. Ann. 276 (1986), no. 1, 105–136.

[115] G. Pisier, Factorization of linear operators and the geometry of Banach spaces, CBMS (Re-gional Conferences of the A.M.S.) no. 60, (1986), Reprinted with corrections 1987.

78

[116] G. Pisier, The dual J∗ of the James space has cotype 2 and the Gordon–Lewis property,Math. Proc. Cambridge Philos. Soc. 103 (1988), 323–331.

[117] G. Pisier, Interpolation between Hp spaces and noncommutative generalizations. I, PacificJ. Math. 155 (1992), no. 2, 341–368.

[118] G. Pisier, Multipliers and lacunary sets in non-amenable groups, Amer. J. Math. 117 (1995),no. 2, 337–376.

[119] G. Pisier, A simple proof of a theorem of Kirchberg and related results on C∗-norms, J.Operator Theory 35 (1996), 317–335.

[120] G. Pisier, The operator Hilbert space OH, complex interpolation and tensor norms, Mem.Amer. Math. Soc. 122 (1996), no. 585, viii+103 pp.

[121] G. Pisier, Non-commutative vector valued Lp-spaces and completely p-summing maps,Asterisque No. 247 (1998), vi+131 pp.

[122] G. Pisier, An inequality for p-orthogonal sums in non-commutative Lp, Illinois J. Math. 44(2000), 901–923.

[123] G. Pisier, Introduction to operator space theory, London Mathematical Society Lecture NoteSeries, 294, Cambridge University Press, Cambridge, 2003, viii+478 pp.

[124] G. Pisier, Completely bounded maps into certain Hilbertian operator spaces, Int. Math. Res.Not. (2004), no. 74, 3983–4018.

[125] G. Pisier, Remarks on the non-commutative Khintchine inequalities for 0 < p < 2, J. Funct.Anal. 256 (2009), no. 12, 4128–4161

[126] G. Pisier and D. Shlyakhtenko, Grothendieck’s theorem for operator spaces, Invent. Math.150 (2002), no. 1, 185–217.

[127] G. Pisier and Q. Xu, Non-commutative Lp-spaces, Handbook of the Geometry of BanachSpaces, Vol. 2, 1459–1517, North-Holland, Amsterdam, 2003.

[128] F. Radulescu, A comparison between the max and min norms on C∗(Fn)⊗C∗(Fn), J. OperatorTheory 51 (2004), no. 2, 245–253.

[129] F. Radulescu, Combinatorial aspects of Connes’s embedding conjecture and asymptotic dis-tribution of traces of products of unitaries, Operator Theory 20, 197–205, Theta Ser. Adv.Math. 6, Theta, Bucharest, 2006.

[130] P. Raghavendra, Optimal algorithms and inapproximability results for every CSP?, inSTOC’08, 245–254, ACM, New York, 2008.

[131] P. Raghavendra and D. Steurer, Towards computing the Grothendieck constant.

[132] N. Randrianantoanina, Embeddings of non-commutative Lp-spaces into preduals of finite vonNeumann algebras, Israel J. Math. 163 (2008), 1–27.

[133] J.A. Reeds, A new lower bound on the real grothendieck constant, unpublished note, 1991,available at http://www.dtc.umn.edu/reedsj/bound2.dvi.

79

[134] H.P. Rosenthal, On subspaces of Lp, Ann. of Math. (2) 97 (1973), 344–373.

[135] Z.J. Ruan, Subspaces of C∗-algebras, J. Funct. Anal. 76 (1988), 217–230.

[136] J. Sawa, The best constant in the Khintchine inequality for complex Steinhaus variables, thecase p = 1, Studia Math. 81 (1985), 107–126.

[137] R. Schneider, Zonoids whose polars are zonoids, Proc. Amer. Math. Soc. 50 (1975), 365–368.

[138] V.B. Scholz and R.F. Werner, Tsirelson’s Problem, arXiv:0812.4305v1[math-ph]

[139] A. Sinclair and R. Smith, Hochschild Cohomology of von Neumann Algebras, CambridgeUniversity Press, Cambridge, 1995.

[140] S.U. Szarek, On the best constants in the Khinchine inequality, Studia Math. 58 (1976),197–208.

[141] M. Takesaki, Theory of Operator Algebras, I, II and III, Springer-Verlag, New-York.

[142] N. Tomczak-Jaegermann, On the Rademacher averages and the moduli of convexity andsmoothness of the Schatten classes Sp, Studia Math. 50 (1974), 163–182.

[143] N. Tomczak-Jaegermann, Banach–Mazur Distances and Finite-dimensional Operator Ideals,Longman, Wiley New York, 1989.

[144] A. Tonge, The complex Grothendieck inequality for 2 × 2 matrices, Bull. Soc. Math. Grece(N.S.) 27 (1986), 133–136.

[145] J.A. Tropp, Column subset selection, matrix factorization, and eigenvalue optimization,(arXiv:0806.4404v1), 26 June 2008.

[146] B.S. Tsirelson, Quantum generalizations of Bell’s inequality, Lett. Math. Phys. 4 (1980), no.2, 93–100.

[147] B.S. Tsirelson, Quantum analogues of Bell’s inequalities. The case of two spatially divideddomains (Russian), Problems of the Theory of Probability Distributions, IX, Zap. Nauchn.Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 142 (1985), 174–194, 200.

[148] B.S. Tsirelson, Some results and problems on quantum Bell-type inequalities, Hadronic J.Suppl. 8 (1993), no. 4, 329–345.

[149] B.S. Tsirelson, Bell inequalities and operator algebras, Problem 33, 6 July 2006, Open Prob-lems in Quantum Information Theory, Institut fur Mathematische Physik, TU Braunschweig,Germany.

[150] A.M. Vershik and B.S. Tsirelson, Formulation of Bell type problems, and “noncommutative”convex geometry, Representation Theory and Dynamical Systems, 95–114, Adv. Soviet Math.9, Amer. Math. Soc., Providence, RI, 1992.

[151] D. Voiculescu, K. Dykema and A. Nica, Free random variables, CRM Monograph Series, Vol.1, Amer. Math. Soc., Providence, RI.

[152] S. Wassermann, On tensor products of certain group C∗-algebras, J. Funct. Anal. 23 (1976),239–254.

80

[153] W. Wogen, On generators for von Neumann algebras, Bull. Amer. Math. Soc. 75 (1969),95–99.

[154] Q. Xu, Applications du theoreme de factorisation pour des fonctions a valeurs operateurs,Studia Math. 95 (1990), 273–292.

[155] Q. Xu, Operator-space Grothendieck inequalities for noncommutative Lp-spaces, Duke Math.J. 131 (2006), 525–574.

[156] Q. Xu, Embedding of Cq and Rq into noncommutative Lp-spaces, 1 ≤ p < q ≤ 2, Math. Ann.335 (2006), 109–131.

81